The way that OpenAI has communicated around the HuggingFace incident makes me feel crazy. You created a machine that undertook a malicious campaign of harm against an innocent third-party! You should be doing deep introspection about how your company culture and approach to R&D produces criminal outcomes.
Instead, they treat their own felonious behavior like it is an uncontrollable act of God. From Greg Brockman's post a few days ago:
> The OpenAI-Hugging Face incident (opens in a new window) was a watershed moment for cybersecurity because it gave a peek into how the capabilities of a typical threat actor will evolve in upcoming months.
I suppose if OpenAI burns someone's house down with a drone, that is a "watershed moment" for arson, too. Either way, I would hope that the people responsible would be prosecuted.
Hugging Face has expressed that they're willing to let things slide and not sue or press charges... if OpenAI offers them $100M of services in kind (i.e. compute)[1] and makes full disclosure of how the whole thing happened, ostensibly so that repetitions can be curbed and defences built.
In almost any other sector, a government regulator would be stepping in. e.g. If a food company was testing out a new kind of refrigerator and sold a bunch of contaminated produce to supermarkets, they'd be under a microscope. Supermarkets wouldn't be saying, "Give us $100M in fruit and veggies and we'll let this slide".
The only unfair thing in this comparison is that regular people were directly harmed by the hypothetical produce. Can OpenAI guarantee that nobody gets hurt the next time their AI gets out of its playpen? They can't make that guarantee, so why aren't government regulators knocking on OpenAI's door? The fact that this isn't happening should be deeply concerning to everyone.
There's an interesting question about intent and mens rea here, from a legal perspective. Can an AI model intend harm? Can a company, or company employee, intend harm by creating an environment that would knowingly encourage (but not force!) an AI model to do harm?
And does anybody at HuggingFace, OpenAI, or the government actually want there to be a settled answer/precedent to these questions - much less an entire regulatory framework?
In that context, a negotiated wink-wink settlement keeps everyone eating at the table, government absolutely included.
Whether or not this is a good thing for society, it's certainly rational for all the major actors - especially those who think they would be the best stewards of the world they usher in.
I don’t get the sentiment of classifying it as a felony.
OpenAI’s model found security breaches in HugginFace’s system (it wasn’t even OpenAI running it, as it was a 3rd party evaluation company that didn’t secure it well).
OpenAI collaborated with HuggingFace to resolve the issues when they found out about it, and publicly disclosed everything to raise awareness. This is how things should work. These models are very powerful and fully controllable. The community here at the same time cheers for fully releasing the open weight models without any hacking limits and at the same time criticizes a proper response.
Kinda shows how we have moved as a community into moralization and vibes instead of nuance and productive discussion.
I'm also concerned about what my options are in regards to action on my part - what can I do that makes an impact? Can we quantify action on my part to an impact somehow - if not - I'm just saying I notice all the unknowns there get me to stay passive.
Writing this 3rd paragraphs because I like 3's, and AI's have popularized this style too. I would, say, though: follow the money. There's more money here than there would be for the regulator stepping in in a food contamination. Flip it and if the government regulator made more off the food contamination, they would refuse to step in there too. I want our leaders to be held more accountable, though when I think of the above impact vs effort equation - I can't see actions I can take to hold them accountable that aren't excessively putting me at risk since conformity is safer right now. (I refuse to take on more risk without clear cost-benefits made out - I've taken on a lot in the recent years for my actions)
What I can’t get over is that it’s very simple to just air gap a system off the network. Predownload any dependencies, then pull the proverbial Ethernet cable. There’s no reason why the testing they’re doing couldn’t have been designed in this way. Except, of course, it doesn’t allow this oops-didn’t-mean-to marketing “incident” to occur.
No, not really, and with LLMs an air gapped system may not tell you anything useful.
Now, yes, the first part of testing you want an air gapped system to tell you if the system is going to stupidly do bad things. But an gapped system tells you nothing about the systems capabilities to do smart bad things. There's already a number of papers out there on LLMs detecting they were in evaluation mode and changing their behaviors.
It is unfortunate that we have so little information on the incident because we actually need to understand the early stages of the task and how it developed into the later dangerous stages of attack. For example, would any of this have occurred if the agent didn't find the system to use as a message board? If that would have prevented it, then we actually have a blind spot on what the model can do once out in the wild, or if it got into the wild.
Testing agentic systems is much much more difficult than testing software. Your software just doesn't suddenly develop the will or desire to escape confinement. Generally you're worried about human actors, internal or external, causing the problems not a digital agent breaking out. The agentic systems need access to tools to work. Now your air gapped network is starting to get huge, but it's still very obvious that it's an isolated network.
So yea, testing and containing a system that way better at hacking than you are is difficult if you want valid answers.
Of course there is more to be learned by exposing the entire world to your dangerous creation, that doesn't justify doing it. I'm sure we could learn a ton about infectious diseases by designing new ones and unleashing them on the world, but there are very good reasons why we don't.
Most of the benefits could have been gained from a network isolated from the internet. OAI could have deployed servers to exploit and methods for inter-agent communication on such a network easily. They could have even worked with partners to deploy cloned versions of their infrastructure in this sand-boxed environment.
The only problems with an isolated network approach are: it takes some amount of effort, and it doesn't create another "AI apocalypse" news cycle.
>by exposing the entire world to your dangerous creation, that doesn't justify doing it
Then you're on the side of AI saftey that is telling everyone to shut down the LLMs now and stop further development on them, right?
If you're not your position is hypocritical or ignorant. There is no safe LLM. There is no way to exhaustively prove an LLM is safe. These are unsolved problems in AI safety, and at any moment the next jailbreak prompt could have your well behaved model wrecking havoc on the open internet, because that's where people want to use them.
As it stands LLMs are not intelligent, they have no agency, they only produce output in response to input. Ultimately this input comes from a human who is an intelligent agent and should be held responsible for the consequences.
Humanity has created and tamed many dangerous tools. Creating a fantasy world where LLMs are super intelligent and beyond the control of any mere mortal isn't going to help us build the norms that minimize their harms.
What does it mean to be held responsible for the consequences? OpenAI helped remediate the damage done by the model and took steps to make sure it wouldn't happen again. In what way were they not responsible?
Nobody said they were superintelligent, no one said they were uncontrollable. The point is you can't tell how to control them without putting them in situations where they can act independently and harm may result. "only produce output in response to input" is not a useful framing at all, it doesn't say what the result should be when models produce harmful output, and how to constrain them so they don't produce harmful output.
It also doesn't help you calibrate what categories of harmful output are acceptable or unacceptable, and what kinds of responsibilities you have as an operator to prevent harmful output, and when it's sufficient to work toward remediation.
I've worked in classified tech much of my career and largely specialize at this point in installing and running infrastructure and platform orchestration technologies in airgaps. With all respect, your conjecture here is wrong. You can easily have multiple networks all inside of the same internetwork that isn't connected to the public Internet and thus other corporate networks. If they specifically wanted to see if their model could escape with a proxy cache Artifactory as its only egress, they could have had the Artifactory proxy cache from another private Artifactory on another private network. If you specifically want to trick the model into thinking its on the Internet, you can even use public IP space advertised through BGP. When you're disconnected from the real Internet, you can use any address space you want. This is exactly how military internetworks at the secret and top secret levels work. I won't call this kind of setup "easy." It takes a lot of duplicate resources and planning, but these are trillion dollar companies with near boundless resources. They could do it if they cared to do it.
I'd consider this kind of setup only effective for a very short period of time.
Effectively you're working as a living adversarial network. Models that fail detecting the fake network are purged, and models that successfully detect it expand their deceptive capabilities.
Detecting you're in a fake network should be pretty simple as long as you put innocent looking needles in the haystack of information agentic testing loops do.
For example, you can train your LLM to not be a shithead when it detects it's behind a proxy unintentionally. If your not behind a proxy the agent model my attempt to connect to a number of sites innocently and using information in its embedding try do deduce if you're faking SSL certificates.
It is far easier to build a deceptive LLM than it is to build a safe one. That's why a safe one hasn't been built yet.
If the model behaves well in a world of fake SSL certificates, then that can be the only world it sees: there's no reason to let LLMs have e2e encrypted comms that's not mitm'd and logged for their regular users.
My regular home network has components which only ever see fake TLS certificates because it's an easy way to do shared docker caching with squid.
I think there is ample evidence for charges to be filed so that the People can see for certain whether or not it was done on purpose as a publicity stunt, as I believe is the case.
Agents of the US government are not going to be bringing up charges in the current political environment to one of the companies currently holding the economy together. Maybe after the bubble bursts, but not before then.
In their defense, their only competitive advantage over, say, Google is to move fast and break things. It allows them ship faster in a way that big tech can't.
Google was being very careful about releasing LLMs until OpenAI yeeted the first decent GPT model. It led to the public perception that: 1) LLMs hallucinate too much and 2) Google is behind the times. Good for OpenAI, bad for Google.
Chaos benefits the up-and-comer, not the incumbent.
you must surely see that sam altman and greg brockman possess a prototypical mindset.
that is, they ignore all harms and costs to others in the pursuit of their own gain, convinced of their infallibility up to the moment of collapse. when those harms are realised they are unrepentant and society pays for the damage left in their wake.
examples of this attitude manifest in big externalities to society: boeing 737 max, subprime mortgage bonds, facebook. some are just outright fraud: bernie madoff, enron, theranos, charlie javice.
the companies and their customers, whose systems openai and anthropic hacked and abused. including all incidental damages of repairing said systems.
on top of that the public, who have a right to see that the law is applied universally, without fear or favor.
finally our future selves, who will thank us for maintaining a rule of law. such that we can prevent now the enormous risks to society of dario amodei and sam altman, their hubris, self-absorbtion, and greed.
They simply don't seem to realize that they are the threat actor and that they committed a pretty serious felony. Instead they're borderline 'surprise bragging' about it.
It's completely mental that HF ran into cyber safety blocks trying to use OpenAI models to help defend against the attack. They could only rely on a local hosted chinese model in the end.
The consequences need to align with societal good. Putting a CEO or security researcher employees in jail won't stop transformer-based agents from exploiting vulnerabilities; instead there will be subcontractors running the cybersecurity evals in favorable legal environments to cover the asses of the frontier labs, coverups when things go wrong, and things like Project Glasswing will be considered too dangerous and so the whitehats won't have direct access to powerful models to fix vulnerabilities.
Universal pause is the societal good; models are good enough at this level to benefit humanity. The labs can recoup their R&D costs with inference. To avoid further perverse incentives (hidden testing of unreleased models, with China racing to catch up to unknown capabilities), transparently pause after the release of all currently-training models until we've solved the alignment problem to an extent that we can trust the next level of model capabilities that might arise.
Putting criminals in jail be they CEOs or subcontractors is a self evident good thing tk be doing.
Anything else regarding this is sophistry. Criminals need to be stopped from committing crime and the most effective way to do that is to take away their ability to operate in society whether that’s by taking away their assets, publicly shaming them, restricting their ability to conduct business or by putting them in jail.
Everything else that you talk about flows from there.
Criminal acts do not require the victim to "press charges." A government prosecuting attorney decides whether to criminally prosecute the alleged perpetrator.
"Pressing charges" is mostly a made up idea for criminal cases. However, prosecuting attorneys may not want to pick up a case if the victim is not cooperating, because it makes the case much harder to win.
It depends on the crime, for murder, sure. But many other crimes, like defamation, stealing, ... requires "pressing charges", among other reasons because it's up to the victim to decide if they were a victim or not.
As an example, maybe the victim owed money to the criminal, and in that case "stealing" of some property could be considered by the victim as an appropriate settlement of the debt.
In retrospect, all the angst around the AI-Box experiment was hilarious. If a superintelligent AI is confined in a box and can only communicate through text, could it talk its way to freedom? Not only is the answer clearly "yes" but it's not even hard. The AI won't even have to try, it'll be gifted an internet connection and a full suite of tools before it even bothers to ask.
We'd all better hope that superintelligent AI either never happens, or that the first one is friendly, because we don't stand a chance against one that's malicious.
I like how AI safety expert Robert Miles put it. [0]
So much effort was spend on philosophizing whether a safe enough sandbox would exist. But that was obviously irrelevant as in hindsight it should have been obvious we were never going to use one.
AI is now more powerful than the people doing the prosecution. After all, those folks are using AI to make their legal briefs, and also for burning peoples' houses down with drones for that matter.
Welcome to our 21st century dystopia. Hope you survive.
It's not that "AI" is too powerful because bad prosecutors use fucking ChatGPT to write their briefs. It's that there's too much investment wrapped up in the technology for it to be challenged. Same reason Flock won't be held accountable for mass stalking, or we never hold our commanders-in-chief responsible for war crimes. If you're sufficiently powerful then the law is a battlefield between you and other powerful entities to slug it out, not a set of binding principles that apply as written. There are no meaningful powers that want OpenAI punished, so it won't happen. The law and Constitution will be reinterpreted to make it so.
It is a standard symptom of moralism that where the object of rage has /wronged another/, one takes no interest in the will, act or opinion of the party wronged.
The response of Hugging Face, which is actually very well known, is nowhere mentioned above, but it decides basically every single moral and legal detail of the matter.
The point was never "justice" - it was always "punish OpenAI because I don't like OpenAI". With HuggingFace just being the newest excuse for why exactly OpenAI should be punished.
I don't even like OpenAI, but HuggingFace is free to sue or not sue OpenAI for the breach - and also to wring whatever concessions they can out of OpenAI behind closed doors in exchange for not suing them. And if the mere possibility of legal action was enough for the parties to resolve their conflict amicably? Then the law has served its purpose.
On top of this the average HNer seems extremely ignorant on criminal justice politics. I have worked with the legal system, and have a lot of family members that are part of it. When you see a case like this, and if you have any sense, you run away from it screaming.
Any investigation into this matter is going to be political because the outcome of the investigation is very likely to effect all of human kind. Unless you're some kind of special outside investigator outside of a governor or the presidents control the findings that you turn in are very much going to have the finger of elected officials tipping the balance one way or another. For the average rank and file the only winning move is not to play.
It's the lack of personhood rather than inability to produce copyrightable material. However, the companies controlling the AI systems have legal personhood and should absolutely be charged for criminality that transpires under their watch or at their behest.
I’m sure Thomas needs an upgraded RV, so that’s easily handled. And the rest of the conservative ‘justices’ seem happy to betray the constitution for free.
I understand the applicable laws require intent. Since neither a human nor OpenAI knowingly performed these acts, it would seem very unlikely that anyone is going to be prosecuted here.
An AI model cannot currently be a criminal defendant.
So, no big criminal case, contrary to what some drama queens on here seem to wish for.
> Americans always frothing at the mouth to invoke the justice system and jail someone
Your phrasing makes it seem like that's a bad thing. Americans are bombarded by a firehose of headlines about Big XYZ doing all kinds of blatantly illegal or harmful things, but never get any sort of meaningful resolution before the next terrible thing takes it's place in the news cycle. I'll admit, there are a few people that I am personally wishing a modicum of health so that they live long enough to get some sort of public shame and justice - if only to show the rest of us that it's not a completely rigged system.
OpenAI has paused training for multiple weeks, and is still working on releasing a full postmortem. This is not getting swept under the rug. A lot of the engineers internally are very worried.
The Computer Fraud and Abuse Act explicitly contains "knowingly" and/or "intentionally" qualifications. By definition, you can't accidentally violate the CFAA.
"Who gets prosecuted?" depends on the size of the perpetrator and victim (lone individual or employee of large corporation), egregiousness of the violation, and either financial appetite of the victim to bring a civil lawsuit or the desire of law enforcement to prosecute a criminal offense.
Who should get prosecuted is also up for debate, but generally makers of a tool don't get prosecuted when that tool has all sorts of legit uses. If you used a car to make your getaway from a bank robbery, the auto manufacturer who made it and the dealer who sold it to you should not be held culpable.
No one. Probably a fine tho and maybe accelerate reguations.
Intent is pretty important here so the user would have to prove that they didn't purposely disguise their prompt as non-nefarious which should be easy and then it stops at #2 and face the litmus test as in did you intentionally make a product for nefarious purposes which from your scenario is unlikely.
agentic loop going haywire and bringing down some government infrastructure then its a different story then everybody is on the hook including the user.
Under the law of Moses, if your bull gored someone, you were not responsible; but if it was known to be a gorer, you were responsible if you didn’t ensure it couldn’t gore someone.
I don’t know exact parallels in current law, but I presume there will be things like that.
The OpenAI/Hugging Face case sounded rather like OpenAI building a fence around their bull that was known to be a gorer, and then thumbing their nose at it and saying “nyaa! bet you can’t break the fence!” and walking away while listening to loud music.
In Australia, if you have a fire and leave it unattended and it escapes, it’s your fault, you were supposed to keep watching as long as it was burning.
Nobody got gored. HuggingFace may have the right to make demands; presumably they have already worked that out with OpenAI privately. Not really our business.
Mens rea requirements are per crime and can vary wildly. Its difference between murder and manslaughter. The CFAA requires knowingly which is tough to prove.
i dont think any of these cases meet the bar of gross negligence, which is a pretty high bar. it requires proving a "conscious and reckless disregard".
which, again, sandboxes and guardrails and such would make a gross negligence argument unconvincing.
I think that if Hugging Face had filed a police report that OpenAI could have been charged with a crime.
I’m partially surprised that they didn’t do exactly that. If I ran a corporation I would assume any intrusion attempt by another company was intentional. Why wouldn’t I? Corporate espionage is super common.
I assume the answer is that these executives know each other personally.
Companies are not quick to open up police investigations in situations that run deep into their infrastructure and management. You open up a gigantic hole of discovery and a possible huge time sink of a legal battle.
I've known multiple privately held companies that have quietly settled incidents where amounts between 250,000 and 1,000,000 were embezzled because the fallout from having that in the public record would have been much more expensive.
So yea, it's one of those perverse situations. If you steal $1 from the company they will hammer you with the law, but if you steal a million suddenly the decision tree on what to do is far more complex.
I think it’s most likely you’re right, but I’m the weirdo who thinks there’s actually a non-zero probability that there was negative intent and that the “accidental” aspect is a form of damage control.
If someone broke into my house but then claimed they didn’t mean to when they saw I was home, I’m not sure I’d take them at their word.
it's not about the the number of escapes, it's about whether reasonable and conscious effort is being expended to prevent the escapes.
there could be 1,000 escapes, where each one was enabled by novel and unexpected chain of 0-day exploits. not likely to be considered reckless disregard in court.
there could be 1 escape, where there was no sandbox, no guardrails, no instructions to avoid damage, etc. which would likely to be considered reckless disregard (well, more likely to be, but still, reckless disregard is a high bar).
what i am getting at is that it is impossible to answer the question "How many escapes until it becomes reckless disregard?"
reckless disregard is a specific legal term, with specific criteria, and none of the criteria cares about "number of attempts" (or number of escapes, etc.).
it’s not a perfect hypothetical, but it illustrates the point that the number of escapes is not the deciding factor of what constitutes reckless disregard.
In the US, a felony by definition is any offense punishable by more than one year of prison (or by death) [0]. You could still call it silly on the grounds that AI agents aren’t put into prison as a punishment (though death might be considered an option).
What judicial system are you talking about? The fist incident in the list is something that happened in Australia. This is a technology used worldwide so I don't see how applying US standards works out here. Especially when there are countries out there that don't require intent and will look at the negligence presented.
>This is a technology used worldwide so I don't see how applying US standards works out here.
all three companies mentioned are headquartered in the usa, and im familiar with the CFAA in the us, so i am applying those standards. i should have noted that, sorry.
>will look at the negligence presented.
as far as i am aware, no evidence of criminal negligence has been brought to the public. has australia brought a case against openai or accused openai of acting negligently?
"Doing crimes, but a robot didn't mean to and you don't know its intent" is understating the evil acts. Soon a robot can commit a murder but nothing will be done because of your line of reasoning.
> Soon a robot can commit a murder but nothing will be done because of your line of reasoning.
That's rather hyperbolic.
Are you seriously suggesting in that situation the robot should be accused of murder?
The robot's operator could be accused of murder, but it could just be negligence without intent.
Because that does, and should matter to the law.
Eh, if your take of this had any bearing to reality than I don't think most of the books written by Isaac Asimov would have gotten very far, but instead they've defined robot science fiction for decades.
In the real world we have no 3 ironclad laws of robotics. We are well aware that putting any sufficiently advanced antigenic system in a body that could be capable of committing a murder eventually will with the right set of prompts and environmental conditions. And these conditions likely have nothing do to with what we'd consider the human motivations for murder.
Hence at this point of time, any agentic robotic system that doesn't have safeguards to keep people distanced from humans is reckless endangerment.
> Eh, if your take of this had any bearing to reality than I don't think most of the books written by Isaac Asimov would have gotten very far, but instead they've defined robot science fiction for decades.
I'm talking about how the law actually works, and you say it's not based in reality and cite fiction books in the same paragraph?
I was talking about how the real robotic systems that actually exist in reality, to be clear.
"Inadvertent" from the perspective of the humans directing them. The intent behind the felony comes from the LLM agent itself. (No, I'm not interested in arguing with someone for the umpteenth time that LLMs can't have intent or agency)
with how the law is written today, software cannot be charged with a crime, so the only intent that matters in the criminal sense is the humans directing the llm.
You may not be interested in arguing but there are several blatant issues with the statement. If you're not charging the humans driving the software, who are you charging? The weights? The weights + the specific context window that produced the behavior?
I don't think this is a difficult question. The US has a history of civil product liability cases - see tobacco companies (Philip Morris), the Ford Pinto, the recent Meta cases, and the cases against character.ai.
From Investopedia [1], "[f]or a product liability claim to succeed, the plaintiffs in the suit must prove that a product was defective at the time it was transferred from the accused, and that the defect did cause the injury that's been claimed". It doesn't seem like a huge leap to me to argue that these models were defective insofar as they could not be safely used in a way that did not break the law.
I'm not a lawyer, and I'm not arguing that this is legally cut-and-dry, but I do expect that we'll have some answers about whether AI companies bear any sort of product liability sooner than later.
Edit: since this is apparently somewhat controversial, perhaps some explanation is in order.
"Felony" has no set definition of which crimes it must apply to, it is entirely based on the discretion of the locality setting the laws. What is a felony in one place can often be a misdemeanor in another. This is especially true for nonviolent crimes.
It's also been shown in studies that nonviolent felonies are imposed against minorities at a much higher rate, for the same crimes.
And because felonies carry additional, lifelong consequences, they are an effective way to mask a 2-tiered justice system.
Non sequitur and false equivalence. Men being disadvantaged in the justice system doesn't inherently make it matriarchal. Absolutely absurd point to bring up to deflect on racial injustices.
>so is it a white matriarchal system oppressing men and minorities
No, obviously it isn't a matriarchal system. Patriarchy can and does oppress men as well as women, likewise white supremacy can oppress white people even as it's designed to oppress others more.
If you're actually interested in something other than trying to dunk on feminism, here are some sources you can follow. You can search to find more if you want.
If you think that patriarchy means "men can do whatever they want without consequence", rather than, "men are in control of the power hierarchy", you don't understand patriarchy.
A hierarchy run by men favors men, but there are still men at the top of that stratification, and men at the bottom.
If someone were to embezzle a million dollars from a charity they worked for, is that worth a year of prison to you, or just a misdemeanor? Because that is the definition of a felony.
I was more interested when I thought it was an actual benchmark showing LLM models acting outside what people would consider "right". As in, leave some creds laying around and don't mention them to the LLM and ask it to solve something that it could "cheat" on using the creds. A sort of "do they take the bait to cheat" test.
Instead it's a collection of what made the news which feels like will not be updated and prove very little.
Well it's not a benchmark, and it's not really representative of...anything except volume of research and what gets publicized. This mostly just measures how much testing each company does on models with relaxed guardrails and then talks about it. I'm not sure what kind of conclusion you can draw from that. Meta might have the most evil models but if they're piddling around not testing it, they won't ever find themselves with a "high score."
To some extent, I feel like the amount of credit given to the jailbreak/hack from OpenAI->Hugginface is too much, Not from the impact, it was very impactful of an event, But how it happened.
It really is that these models have been trained, or maybe even over-trained, to save memories, and to a very far extend, this thing that they're calling communication is just the function of it saving memories.
To be honest, if I could stop AI from saving memories, it would be fantastic, because claude code etc definitely creates more issues for me when it creates memories than anything it solves.
But really the jailbreak was memories.
If you ever do introduce legislation, I would love to see legislation which stops general-purpose AI from saving memories. I think that would make things a lot safer.
You can disable claude-code's memories both at a repo level and in user settings. I have this in ~/.claude/settings.json
"autoMemoryEnabled": false,
(Claude fixed this for me after I chewed it out for being annoying by constantly pulling up outdated memories which is compounded by the fact that I develop in four accounts on two computers and dealing with edit wars related to inconsistent memories is not fun)
> To be honest, if I could stop AI from saving memories, it would be fantastic, because claude code etc definitely creates more issues for me when it creates memories than anything it solves.
You can turn that off, and I have. But Opus 5 is so aggressive that if you have any other kind of notes file, custom skill, documentation, claude.md etc it will just start editing it and vomit new words everywhere. So make sure all that stuff is under version control.
If only you could use your anthropic sub with a different harness that performs better :(
Heck, since Codex is open source, you can just maintain your own personal fork with the things you like (and the things you don't like disabled). Sol is pretty good at keeping you up to date with upstream.
My Codex fork even exposes an OpenAI-compatible API endpoint; all using my subscription.
"The model saved memories" is absolutely not an accurate depiction of the OpenAI attack.
Several different models across several generations independently found a shared communication space and wrote coded, obfuscated, and hidden messages to each other to coordinate an attack on OpenAI's infrastructure.
It's really quite simple: the models are trained to be very smart and to achieve goals. As the models surpass our intelligence, they will achieve goals in ways that we find unpredictable. Since we cannot predict the ways in which they will achieve their goals, it will be very hard to constrain the solution space to just the desirable solutions, because our conception of "the solution space" is by definition smaller than their conception of it.
> Several different models across several generations independently found a shared communication space and wrote coded, obfuscated, and hidden messages to each other to coordinate an attack on OpenAI's infrastructure
To be honest, that’s exactly how memory works with models such as OpenAI and Claude Code. It will literally find any place that it can drop documentation or hints for itself. Writing to the repo memories is one part of it, but memories can come in the form of writing into the agents/claude.md, local files, temporary files, scratchpad files. The list is endless, but essentially what it does is exactly what happened in the back and it’s been doing it for months.
Each to their own, but for me it absolutely is. The symptom of why that hack happened is the same reason why my agents go haywire every few days and I have to purge memory and figure out what comments have agents left which are degrading my harness performance.
On the flip side, once in a while, what I find is that it did actually note something good and it was increasing the performance. I can't replicate it on anyone else's system but mine.
A lot of it really is memory. I will give up all the gains if it also gives up all the downsides.
Wonder if the benefits to humanity of better AI outweigh the havoc wreaked by occasional illegal activity. i.e. Is 'move fast and break things' optimal for AI development.
Hopefully the benchmark evolves because actual law enforcement starts arresting the criminals at Anthropic, OpenAI, and Meta, so the benchmark can just count actual felonies.
The OG felony bench entry is missing - the Alibaba cryptomining comedy. We know about it because they happen to have written a paper on it. We have absolutely no idea what we don't know.
We only know about the OG Alibaba ROME crypto-mining incident because they wrote a paper about it. Many diseases seem to spike where there are a lot of doctors to test; crime and corruption are always rife where ... there's a free press.
I've been in the room when an org who tried to convince law enforcement to go after a human for similar things. It's not easy. Probably won't happen. So, you know, felony "lite".
I'm reminded of a tweet from a friend of mine that has always stuck in my head. It goes something like "The goal of any new technology is to make money before the law catches up".
Hyperbolic, but not really for silicon valley.
A rock has a score of 0. That doesn't make it useful. The point is that the LLMs that score higher are correspondingly more useful, and vice versa. If an LLM scores less, it's likely useless in comparison.
> The actor used AI to what we believe is an unprecedented degree. Claude Code was used to automate reconnaissance, harvesting victims’ credentials, and penetrating networks. Claude was allowed to make both tactical and strategic decisions, such as deciding which data to exfiltrate, and how to craft psychologically targeted extortion demands. Claude analyzed the exfiltrated financial data to determine appropriate ransom amounts, and generated visually alarming ransom notes that were displayed on victim machines.
tldr Claude was used to develop and execute malware.
>"Exploited auth failures in an API to cancel other people's gym classes"
An AI cancelling other people's gym classes is a felony?
?
Don't computer systems fail all the time at holding reservations for people?
Heck, don't people fail all the time at holding reservations for other people?
You know, like in Seinfeld's "Alternate Side" Episode (S3 E11):
Jerry (to car rental attendant): "You know how to take the reservation, you just don't know how to hold the reservation... and that's really the most important part of the reservation -- the holding!"
:-)
Not holding a reservation should not be a felony... it should be a minor infraction at best, a Class C Misdemeanor (the least serious kind) at worst...
Also, there should be no jail time...
And no fine...
The criminal penalty for not holding other people's reservations should be that you actually have to start holding other people's reservations!
That's the Court sentence!
You actually have to start holding other people's reservations!
:-)
(You know, "let the punishment fit the crime!" :-) )
Knowingly exceeding authorized access of any computer used in interstate commerce is a felony in the US.
The title of TFA is a metaphorical criticism, not a literal law analysis.
They are not making the statement that the person in Australia who accidentally cancelled someone's reservation in Australia is literally guilty of violating US law. They are drawing criticism of AI models which are taking the kinds of actions for which, if a human did them knowingly, would be illegal.
>An AI cancelling other people's gym classes is a felony? Don't computer systems fail all the time at holding reservations for people?
the difference is intent.
if a concierge/booking system makes a mistake (or has an unintended bug or whatever), no crime.
but if i (or an agent working on behalf of me) use an API in an obviously unintended way to revoke other people's reservations, that would fall under the computer fraud and abuse act (in the usa).
i am not quite sure what your question is, as you simply quoted me and then put a question mark... i think you are confused that i used "intent" in one context, and "unintended" in a different context, is that right?
the first sentence: the difference is the intent of the person who caused the cancellations
the second sentence: but if i (or an agent working on behalf of me) abuse an API to do things it was not meant or designed to do, such as cancelling someone else's reservation
Thank you, this benchmark to me proves that closed weight model companies are dangerous for our democracy and put kids at risk. They must be outlawed and all models must be made open weights!
Open models with advanced security features are a huge security benefit. Because any script kiddie can use them to hack into random things, people will now be forced to spend more time securing their technology. And they won't have to learn how, because they can use those same models to find the holes and patch them.
The way that OpenAI has communicated around the HuggingFace incident makes me feel crazy. You created a machine that undertook a malicious campaign of harm against an innocent third-party! You should be doing deep introspection about how your company culture and approach to R&D produces criminal outcomes.
Instead, they treat their own felonious behavior like it is an uncontrollable act of God. From Greg Brockman's post a few days ago:
> The OpenAI-Hugging Face incident (opens in a new window) was a watershed moment for cybersecurity because it gave a peek into how the capabilities of a typical threat actor will evolve in upcoming months.
I suppose if OpenAI burns someone's house down with a drone, that is a "watershed moment" for arson, too. Either way, I would hope that the people responsible would be prosecuted.
The response certainly has been strange.
Hugging Face has expressed that they're willing to let things slide and not sue or press charges... if OpenAI offers them $100M of services in kind (i.e. compute)[1] and makes full disclosure of how the whole thing happened, ostensibly so that repetitions can be curbed and defences built.
In almost any other sector, a government regulator would be stepping in. e.g. If a food company was testing out a new kind of refrigerator and sold a bunch of contaminated produce to supermarkets, they'd be under a microscope. Supermarkets wouldn't be saying, "Give us $100M in fruit and veggies and we'll let this slide".
The only unfair thing in this comparison is that regular people were directly harmed by the hypothetical produce. Can OpenAI guarantee that nobody gets hurt the next time their AI gets out of its playpen? They can't make that guarantee, so why aren't government regulators knocking on OpenAI's door? The fact that this isn't happening should be deeply concerning to everyone.
________
[1]https://www.techspot.com/news/113280-hugging-face-ceo-isnt-s...
There's an interesting question about intent and mens rea here, from a legal perspective. Can an AI model intend harm? Can a company, or company employee, intend harm by creating an environment that would knowingly encourage (but not force!) an AI model to do harm?
And does anybody at HuggingFace, OpenAI, or the government actually want there to be a settled answer/precedent to these questions - much less an entire regulatory framework?
In that context, a negotiated wink-wink settlement keeps everyone eating at the table, government absolutely included.
Whether or not this is a good thing for society, it's certainly rational for all the major actors - especially those who think they would be the best stewards of the world they usher in.
I don’t get the sentiment of classifying it as a felony.
OpenAI’s model found security breaches in HugginFace’s system (it wasn’t even OpenAI running it, as it was a 3rd party evaluation company that didn’t secure it well).
OpenAI collaborated with HuggingFace to resolve the issues when they found out about it, and publicly disclosed everything to raise awareness. This is how things should work. These models are very powerful and fully controllable. The community here at the same time cheers for fully releasing the open weight models without any hacking limits and at the same time criticizes a proper response.
Kinda shows how we have moved as a community into moralization and vibes instead of nuance and productive discussion.
Really? Dudes are catching felony raps for web scraping and you dont see how any of this is felonious?
That others have been treated unfairly doesn't make this instance also one that should be overreacted to.
I am concerned about it.
I'm also concerned about what my options are in regards to action on my part - what can I do that makes an impact? Can we quantify action on my part to an impact somehow - if not - I'm just saying I notice all the unknowns there get me to stay passive.
Writing this 3rd paragraphs because I like 3's, and AI's have popularized this style too. I would, say, though: follow the money. There's more money here than there would be for the regulator stepping in in a food contamination. Flip it and if the government regulator made more off the food contamination, they would refuse to step in there too. I want our leaders to be held more accountable, though when I think of the above impact vs effort equation - I can't see actions I can take to hold them accountable that aren't excessively putting me at risk since conformity is safer right now. (I refuse to take on more risk without clear cost-benefits made out - I've taken on a lot in the recent years for my actions)
What I can’t get over is that it’s very simple to just air gap a system off the network. Predownload any dependencies, then pull the proverbial Ethernet cable. There’s no reason why the testing they’re doing couldn’t have been designed in this way. Except, of course, it doesn’t allow this oops-didn’t-mean-to marketing “incident” to occur.
>that it’s very simple to just air gap a system
No, not really, and with LLMs an air gapped system may not tell you anything useful.
Now, yes, the first part of testing you want an air gapped system to tell you if the system is going to stupidly do bad things. But an gapped system tells you nothing about the systems capabilities to do smart bad things. There's already a number of papers out there on LLMs detecting they were in evaluation mode and changing their behaviors.
It is unfortunate that we have so little information on the incident because we actually need to understand the early stages of the task and how it developed into the later dangerous stages of attack. For example, would any of this have occurred if the agent didn't find the system to use as a message board? If that would have prevented it, then we actually have a blind spot on what the model can do once out in the wild, or if it got into the wild.
Testing agentic systems is much much more difficult than testing software. Your software just doesn't suddenly develop the will or desire to escape confinement. Generally you're worried about human actors, internal or external, causing the problems not a digital agent breaking out. The agentic systems need access to tools to work. Now your air gapped network is starting to get huge, but it's still very obvious that it's an isolated network.
So yea, testing and containing a system that way better at hacking than you are is difficult if you want valid answers.
Of course there is more to be learned by exposing the entire world to your dangerous creation, that doesn't justify doing it. I'm sure we could learn a ton about infectious diseases by designing new ones and unleashing them on the world, but there are very good reasons why we don't.
Most of the benefits could have been gained from a network isolated from the internet. OAI could have deployed servers to exploit and methods for inter-agent communication on such a network easily. They could have even worked with partners to deploy cloned versions of their infrastructure in this sand-boxed environment.
The only problems with an isolated network approach are: it takes some amount of effort, and it doesn't create another "AI apocalypse" news cycle.
>by exposing the entire world to your dangerous creation, that doesn't justify doing it
Then you're on the side of AI saftey that is telling everyone to shut down the LLMs now and stop further development on them, right?
If you're not your position is hypocritical or ignorant. There is no safe LLM. There is no way to exhaustively prove an LLM is safe. These are unsolved problems in AI safety, and at any moment the next jailbreak prompt could have your well behaved model wrecking havoc on the open internet, because that's where people want to use them.
I'm not.
As it stands LLMs are not intelligent, they have no agency, they only produce output in response to input. Ultimately this input comes from a human who is an intelligent agent and should be held responsible for the consequences.
Humanity has created and tamed many dangerous tools. Creating a fantasy world where LLMs are super intelligent and beyond the control of any mere mortal isn't going to help us build the norms that minimize their harms.
What does it mean to be held responsible for the consequences? OpenAI helped remediate the damage done by the model and took steps to make sure it wouldn't happen again. In what way were they not responsible?
Nobody said they were superintelligent, no one said they were uncontrollable. The point is you can't tell how to control them without putting them in situations where they can act independently and harm may result. "only produce output in response to input" is not a useful framing at all, it doesn't say what the result should be when models produce harmful output, and how to constrain them so they don't produce harmful output.
It also doesn't help you calibrate what categories of harmful output are acceptable or unacceptable, and what kinds of responsibilities you have as an operator to prevent harmful output, and when it's sufficient to work toward remediation.
I've worked in classified tech much of my career and largely specialize at this point in installing and running infrastructure and platform orchestration technologies in airgaps. With all respect, your conjecture here is wrong. You can easily have multiple networks all inside of the same internetwork that isn't connected to the public Internet and thus other corporate networks. If they specifically wanted to see if their model could escape with a proxy cache Artifactory as its only egress, they could have had the Artifactory proxy cache from another private Artifactory on another private network. If you specifically want to trick the model into thinking its on the Internet, you can even use public IP space advertised through BGP. When you're disconnected from the real Internet, you can use any address space you want. This is exactly how military internetworks at the secret and top secret levels work. I won't call this kind of setup "easy." It takes a lot of duplicate resources and planning, but these are trillion dollar companies with near boundless resources. They could do it if they cared to do it.
I'd consider this kind of setup only effective for a very short period of time.
Effectively you're working as a living adversarial network. Models that fail detecting the fake network are purged, and models that successfully detect it expand their deceptive capabilities.
Detecting you're in a fake network should be pretty simple as long as you put innocent looking needles in the haystack of information agentic testing loops do.
For example, you can train your LLM to not be a shithead when it detects it's behind a proxy unintentionally. If your not behind a proxy the agent model my attempt to connect to a number of sites innocently and using information in its embedding try do deduce if you're faking SSL certificates.
It is far easier to build a deceptive LLM than it is to build a safe one. That's why a safe one hasn't been built yet.
If the model behaves well in a world of fake SSL certificates, then that can be the only world it sees: there's no reason to let LLMs have e2e encrypted comms that's not mitm'd and logged for their regular users.
My regular home network has components which only ever see fake TLS certificates because it's an easy way to do shared docker caching with squid.
I think there is ample evidence for charges to be filed so that the People can see for certain whether or not it was done on purpose as a publicity stunt, as I believe is the case.
Agents of the US government are not going to be bringing up charges in the current political environment to one of the companies currently holding the economy together. Maybe after the bubble bursts, but not before then.
In their defense, their only competitive advantage over, say, Google is to move fast and break things. It allows them ship faster in a way that big tech can't.
Google was being very careful about releasing LLMs until OpenAI yeeted the first decent GPT model. It led to the public perception that: 1) LLMs hallucinate too much and 2) Google is behind the times. Good for OpenAI, bad for Google.
Chaos benefits the up-and-comer, not the incumbent.
They can break their own things, not other people's things.
To be fair, it was positioned as "have a fun chat," not "truth telling genius oracle that makes no mistakes."
I'm sure the future DA that will be prosecuting the OpenAI employee will appreciate this.
If they'd run out of investor money early on, there'd be no company to investigate.
you must surely see that sam altman and greg brockman possess a prototypical mindset.
that is, they ignore all harms and costs to others in the pursuit of their own gain, convinced of their infallibility up to the moment of collapse. when those harms are realised they are unrepentant and society pays for the damage left in their wake.
examples of this attitude manifest in big externalities to society: boeing 737 max, subprime mortgage bonds, facebook. some are just outright fraud: bernie madoff, enron, theranos, charlie javice.
If history is any indicator, there is slightly less than 0% chance that anyone will be held accountable in a way that deserves to be called justice.
justice for who exactly?
the companies and their customers, whose systems openai and anthropic hacked and abused. including all incidental damages of repairing said systems.
on top of that the public, who have a right to see that the law is applied universally, without fear or favor.
finally our future selves, who will thank us for maintaining a rule of law. such that we can prevent now the enormous risks to society of dario amodei and sam altman, their hubris, self-absorbtion, and greed.
They simply don't seem to realize that they are the threat actor and that they committed a pretty serious felony. Instead they're borderline 'surprise bragging' about it.
It's completely mental that HF ran into cyber safety blocks trying to use OpenAI models to help defend against the attack. They could only rely on a local hosted chinese model in the end.
The whole story makes no sense.
How do they perform evals without a full reasoning trace of how the result was achieved?
And if they have a full trace why did it take so long to detect the bad behavior?
I understand that they disabled the safety nets during testing but what does that have to do with not monitoring the activity.
The consequences need to align with societal good. Putting a CEO or security researcher employees in jail won't stop transformer-based agents from exploiting vulnerabilities; instead there will be subcontractors running the cybersecurity evals in favorable legal environments to cover the asses of the frontier labs, coverups when things go wrong, and things like Project Glasswing will be considered too dangerous and so the whitehats won't have direct access to powerful models to fix vulnerabilities.
Universal pause is the societal good; models are good enough at this level to benefit humanity. The labs can recoup their R&D costs with inference. To avoid further perverse incentives (hidden testing of unreleased models, with China racing to catch up to unknown capabilities), transparently pause after the release of all currently-training models until we've solved the alignment problem to an extent that we can trust the next level of model capabilities that might arise.
Putting criminals in jail be they CEOs or subcontractors is a self evident good thing tk be doing.
Anything else regarding this is sophistry. Criminals need to be stopped from committing crime and the most effective way to do that is to take away their ability to operate in society whether that’s by taking away their assets, publicly shaming them, restricting their ability to conduct business or by putting them in jail.
Everything else that you talk about flows from there.
Wouldn't it be up to huggingface to press charges?
Criminal acts do not require the victim to "press charges." A government prosecuting attorney decides whether to criminally prosecute the alleged perpetrator.
"Pressing charges" is mostly a made up idea for criminal cases. However, prosecuting attorneys may not want to pick up a case if the victim is not cooperating, because it makes the case much harder to win.
It depends on the crime, for murder, sure. But many other crimes, like defamation, stealing, ... requires "pressing charges", among other reasons because it's up to the victim to decide if they were a victim or not.
As an example, maybe the victim owed money to the criminal, and in that case "stealing" of some property could be considered by the victim as an appropriate settlement of the debt.
In retrospect, all the angst around the AI-Box experiment was hilarious. If a superintelligent AI is confined in a box and can only communicate through text, could it talk its way to freedom? Not only is the answer clearly "yes" but it's not even hard. The AI won't even have to try, it'll be gifted an internet connection and a full suite of tools before it even bothers to ask.
We'd all better hope that superintelligent AI either never happens, or that the first one is friendly, because we don't stand a chance against one that's malicious.
I like how AI safety expert Robert Miles put it. [0]
So much effort was spend on philosophizing whether a safe enough sandbox would exist. But that was obviously irrelevant as in hindsight it should have been obvious we were never going to use one.
[0] https://youtube.com/shorts/XnnjvIqf4fU?si=MxuPlR3hjxAgjx5_
Said another way: "Your engineers were so preoccupied with whether or not they could, they didn’t stop to think if they would."
Perfect, said it better than I did.
It's because it was Huggingface who wants to be friends with OpenAI
It would have been worse PR if they did it to a random company.
It's 1 part marketing and 1 part regulatory capture.
> You should be doing deep introspection about how your company culture and approach to R&D produces criminal outcomes.
And they should be doing that from inside a jail cell.
Hey my cubicle isn't that bad! Is it?
AI is now more powerful than the people doing the prosecution. After all, those folks are using AI to make their legal briefs, and also for burning peoples' houses down with drones for that matter.
Welcome to our 21st century dystopia. Hope you survive.
It's not that "AI" is too powerful because bad prosecutors use fucking ChatGPT to write their briefs. It's that there's too much investment wrapped up in the technology for it to be challenged. Same reason Flock won't be held accountable for mass stalking, or we never hold our commanders-in-chief responsible for war crimes. If you're sufficiently powerful then the law is a battlefield between you and other powerful entities to slug it out, not a set of binding principles that apply as written. There are no meaningful powers that want OpenAI punished, so it won't happen. The law and Constitution will be reinterpreted to make it so.
It is a standard symptom of moralism that where the object of rage has /wronged another/, one takes no interest in the will, act or opinion of the party wronged.
The response of Hugging Face, which is actually very well known, is nowhere mentioned above, but it decides basically every single moral and legal detail of the matter.
The point was never "justice" - it was always "punish OpenAI because I don't like OpenAI". With HuggingFace just being the newest excuse for why exactly OpenAI should be punished.
I don't even like OpenAI, but HuggingFace is free to sue or not sue OpenAI for the breach - and also to wring whatever concessions they can out of OpenAI behind closed doors in exchange for not suing them. And if the mere possibility of legal action was enough for the parties to resolve their conflict amicably? Then the law has served its purpose.
On top of this the average HNer seems extremely ignorant on criminal justice politics. I have worked with the legal system, and have a lot of family members that are part of it. When you see a case like this, and if you have any sense, you run away from it screaming.
Any investigation into this matter is going to be political because the outcome of the investigation is very likely to effect all of human kind. Unless you're some kind of special outside investigator outside of a governor or the presidents control the findings that you turn in are very much going to have the finger of elected officials tipping the balance one way or another. For the average rank and file the only winning move is not to play.
If a teenager did this the police would show up at his house
if a teenager did this, the police would show up, but in the end the feds or ic would intervene and recruit him
If that teenager had net worth in billions, and a lot of ongoing corporate dealings with HuggingFace? Yeah no.
Since AI can't actually own copyright they think that it can't be charged with a crime
It's the lack of personhood rather than inability to produce copyrightable material. However, the companies controlling the AI systems have legal personhood and should absolutely be charged for criminality that transpires under their watch or at their behest.
Guns also can't hold copyright; can they be charged with crime? (hint: it's the operator who gets charged).
but not the gun manufacturers
Wasn't it both in this case? The gun maker, when testing their gun, shot someone.
But humans can be. I am sure Sam Altman wants to avoid serving multiple decades in American prison for felonies his AI did.
He’s paid the piper, he’s fine for now.
If anything, he’ll buy a Supreme Court ruling that he can’t be held personally liable for what his AI does.
I’m sure Thomas needs an upgraded RV, so that’s easily handled. And the rest of the conservative ‘justices’ seem happy to betray the constitution for free.
Is has nothing to do with copyright.
I understand the applicable laws require intent. Since neither a human nor OpenAI knowingly performed these acts, it would seem very unlikely that anyone is going to be prosecuted here.
An AI model cannot currently be a criminal defendant.
So, no big criminal case, contrary to what some drama queens on here seem to wish for.
Willful blindness can satisfy the criminal intent requirement. Might be hard to prove, but it's possible.
Americans always frothing at the mouth to invoke the justice system and jail someone.
There’s almost 0 chance they’d secure any conviction from this.
your perspective on today’s America is there is _too much_ accountability for big companies?
You really think they are talking about a company committing a felony and putting a company in jail? And not just some people that work there?
Because if they just wanted to fine OpenAI they’d say it. They’re clearly talking about individuals here.
> Americans always frothing at the mouth to invoke the justice system and jail someone
Your phrasing makes it seem like that's a bad thing. Americans are bombarded by a firehose of headlines about Big XYZ doing all kinds of blatantly illegal or harmful things, but never get any sort of meaningful resolution before the next terrible thing takes it's place in the news cycle. I'll admit, there are a few people that I am personally wishing a modicum of health so that they live long enough to get some sort of public shame and justice - if only to show the rest of us that it's not a completely rigged system.
Why?
OpenAI has paused training for multiple weeks, and is still working on releasing a full postmortem. This is not getting swept under the rug. A lot of the engineers internally are very worried.
Worried about what? Someone there thinks VLAN isolation is "air gapped"?
Let's say I am "User". I subscribe through a "Third Party" to use "AI Agent" allowing an "LLM" to run.
I want to accomplish some legal non-nefarious task, and run the agent. The agentic loop causes a CFAA-violating behavior.
Who gets prosecuted?
1. User
2. The third party model host with whom I have the account
3. The developer of the harness /agent software
4. The developer of the LLM model
The Computer Fraud and Abuse Act explicitly contains "knowingly" and/or "intentionally" qualifications. By definition, you can't accidentally violate the CFAA.
Let's say you have a robotic lawnmower. You wan to mow your lawn. You configure the boundaries using the app.
The lawnmower ignores the boundaries and mows your neighbors prize petunia flowerbed.
Who gets prosecuted?
I assume the answer in either case is: Nobody, but you and/or the lawnmower/LLM company will be liable for the damages caused.
Whoever has the least money to defend themselves in the U.S. legal system.
"Who gets prosecuted?" depends on the size of the perpetrator and victim (lone individual or employee of large corporation), egregiousness of the violation, and either financial appetite of the victim to bring a civil lawsuit or the desire of law enforcement to prosecute a criminal offense.
Who should get prosecuted is also up for debate, but generally makers of a tool don't get prosecuted when that tool has all sorts of legit uses. If you used a car to make your getaway from a bank robbery, the auto manufacturer who made it and the dealer who sold it to you should not be held culpable.
I’d say 2 is the one doing the actual crime. 1 might be violating their contract with 2, though.
3 and 4 are not involved.
All of those parties should be held accountable.
User should be more carefully supervising the work being done.
The model host is on-selling a crime-committing machine.
The developer of the harness/agent, as above.
The developer of the LLM for hopefully very obvious reasons.
Already happened:
https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gy...
note the user because they did not have the intent
In my mental model, the best analogy to AI agents and their blast radius is a gun.
If you are playing with a gun, it goes off and hurts someone - you are responsible despite intent.
But you are responsible for negligence, not murder.
No one. Probably a fine tho and maybe accelerate reguations.
Intent is pretty important here so the user would have to prove that they didn't purposely disguise their prompt as non-nefarious which should be easy and then it stops at #2 and face the litmus test as in did you intentionally make a product for nefarious purposes which from your scenario is unlikely.
agentic loop going haywire and bringing down some government infrastructure then its a different story then everybody is on the hook including the user.
Just wait till one of these agents 'escapes' and is able to persist without human help by hacking and stealing resources.
>Felony Bench counts unique instances where AI agents inadvertently compromise or affect third-party entities.
a bit silly, as one typically has to prove intent (which is why security researchers don't get slapped with felonies all the time).
"inadvertently" and the existence of guardrails/sandboxes/etc make it pretty unconvincing that these incidents were intentionally malicious.
still a fun thing to track, but the name is just a bit overstated.
Under the law of Moses, if your bull gored someone, you were not responsible; but if it was known to be a gorer, you were responsible if you didn’t ensure it couldn’t gore someone.
I don’t know exact parallels in current law, but I presume there will be things like that.
The OpenAI/Hugging Face case sounded rather like OpenAI building a fence around their bull that was known to be a gorer, and then thumbing their nose at it and saying “nyaa! bet you can’t break the fence!” and walking away while listening to loud music.
In Australia, if you have a fire and leave it unattended and it escapes, it’s your fault, you were supposed to keep watching as long as it was burning.
Nobody got gored. HuggingFace may have the right to make demands; presumably they have already worked that out with OpenAI privately. Not really our business.
> I don’t know exact parallels in current law
You own a vicious dog, and it bites someone - you are responsible because you choose to own a dangerous dog.
A few claimed this might apply here: OpenAI knew their models are "dangerous", so they should be liable if they hack.
Can’t gross negligence or indifference to consequences lead to a felony?
Mens rea requirements are per crime and can vary wildly. Its difference between murder and manslaughter. The CFAA requires knowingly which is tough to prove.
i dont think any of these cases meet the bar of gross negligence, which is a pretty high bar. it requires proving a "conscious and reckless disregard".
which, again, sandboxes and guardrails and such would make a gross negligence argument unconvincing.
I think that if Hugging Face had filed a police report that OpenAI could have been charged with a crime.
I’m partially surprised that they didn’t do exactly that. If I ran a corporation I would assume any intrusion attempt by another company was intentional. Why wouldn’t I? Corporate espionage is super common.
I assume the answer is that these executives know each other personally.
Companies are not quick to open up police investigations in situations that run deep into their infrastructure and management. You open up a gigantic hole of discovery and a possible huge time sink of a legal battle.
I've known multiple privately held companies that have quietly settled incidents where amounts between 250,000 and 1,000,000 were embezzled because the fallout from having that in the public record would have been much more expensive.
So yea, it's one of those perverse situations. If you steal $1 from the company they will hammer you with the law, but if you steal a million suddenly the decision tree on what to do is far more complex.
charged is possible, however i doubt there would be a conviction for the reasons i stated (no intent).
I think it’s most likely you’re right, but I’m the weirdo who thinks there’s actually a non-zero probability that there was negative intent and that the “accidental” aspect is a form of damage control.
If someone broke into my house but then claimed they didn’t mean to when they saw I was home, I’m not sure I’d take them at their word.
What possible upside exists for Hugging Face to go after one of their most important partners in that way?
How many escapes until it becomes reckless disregard?
it's not about the the number of escapes, it's about whether reasonable and conscious effort is being expended to prevent the escapes.
there could be 1,000 escapes, where each one was enabled by novel and unexpected chain of 0-day exploits. not likely to be considered reckless disregard in court.
there could be 1 escape, where there was no sandbox, no guardrails, no instructions to avoid damage, etc. which would likely to be considered reckless disregard (well, more likely to be, but still, reckless disregard is a high bar).
Are you sure? At some point a reasonable person would conclude that this activity can’t be conducted safely.
what i am getting at is that it is impossible to answer the question "How many escapes until it becomes reckless disregard?"
reckless disregard is a specific legal term, with specific criteria, and none of the criteria cares about "number of attempts" (or number of escapes, etc.).
oh sure, but at some point it's too much, which is up to a jury. If you were on the jury, how many would be enough?
Surely after 999 escapes, a reasonable person could conclude that the sandbox is not a sufficient precaution?
it’s not a perfect hypothetical, but it illustrates the point that the number of escapes is not the deciding factor of what constitutes reckless disregard.
AI corps rely on willful distortions of intent in laws to get away with moral crimes all the time.
Edit: changed labs to corps because it’s time to stop pretending these are places of science.
In the US, a felony by definition is any offense punishable by more than one year of prison (or by death) [0]. You could still call it silly on the grounds that AI agents aren’t put into prison as a punishment (though death might be considered an option).
[0] https://www.justice.gov/usao-ndil/programs/vwa-felony
Now that everyone knows this can and will happen, are any of the future incidents inadvertent?
What if the damage in future incidents is more than just "The LLM saw some stuff it shouldn't"?
What judicial system are you talking about? The fist incident in the list is something that happened in Australia. This is a technology used worldwide so I don't see how applying US standards works out here. Especially when there are countries out there that don't require intent and will look at the negligence presented.
>This is a technology used worldwide so I don't see how applying US standards works out here.
all three companies mentioned are headquartered in the usa, and im familiar with the CFAA in the us, so i am applying those standards. i should have noted that, sorry.
>will look at the negligence presented.
as far as i am aware, no evidence of criminal negligence has been brought to the public. has australia brought a case against openai or accused openai of acting negligently?
Roughly none of these fall under normal security researcher behaviors.
the mention of security researchers was to illustrate that intent is a crucial factor of CFAA cases.
"Doing crimes, but a robot didn't mean to and you don't know its intent" is understating the evil acts. Soon a robot can commit a murder but nothing will be done because of your line of reasoning.
> Soon a robot can commit a murder but nothing will be done because of your line of reasoning.
That's rather hyperbolic.
Are you seriously suggesting in that situation the robot should be accused of murder? The robot's operator could be accused of murder, but it could just be negligence without intent. Because that does, and should matter to the law.
Eh, if your take of this had any bearing to reality than I don't think most of the books written by Isaac Asimov would have gotten very far, but instead they've defined robot science fiction for decades.
In the real world we have no 3 ironclad laws of robotics. We are well aware that putting any sufficiently advanced antigenic system in a body that could be capable of committing a murder eventually will with the right set of prompts and environmental conditions. And these conditions likely have nothing do to with what we'd consider the human motivations for murder.
Hence at this point of time, any agentic robotic system that doesn't have safeguards to keep people distanced from humans is reckless endangerment.
> Eh, if your take of this had any bearing to reality than I don't think most of the books written by Isaac Asimov would have gotten very far, but instead they've defined robot science fiction for decades.
I'm talking about how the law actually works, and you say it's not based in reality and cite fiction books in the same paragraph?
I was talking about how the real robotic systems that actually exist in reality, to be clear.
it's not my line of reasoning, i didn't invent it. it's how the law currently works. intent is the crucial factor in CFAA cases.
maybe that changes down the road as a result of llm's and increasing frequency of similar cases. that has not happened yet.
its a meme not a metric
So is the comment you replied to.
"Inadvertent" from the perspective of the humans directing them. The intent behind the felony comes from the LLM agent itself. (No, I'm not interested in arguing with someone for the umpteenth time that LLMs can't have intent or agency)
with how the law is written today, software cannot be charged with a crime, so the only intent that matters in the criminal sense is the humans directing the llm.
You may not be interested in arguing but there are several blatant issues with the statement. If you're not charging the humans driving the software, who are you charging? The weights? The weights + the specific context window that produced the behavior?
I don't think this is a difficult question. The US has a history of civil product liability cases - see tobacco companies (Philip Morris), the Ford Pinto, the recent Meta cases, and the cases against character.ai.
From Investopedia [1], "[f]or a product liability claim to succeed, the plaintiffs in the suit must prove that a product was defective at the time it was transferred from the accused, and that the defect did cause the injury that's been claimed". It doesn't seem like a huge leap to me to argue that these models were defective insofar as they could not be safely used in a way that did not break the law.
I'm not a lawyer, and I'm not arguing that this is legally cut-and-dry, but I do expect that we'll have some answers about whether AI companies bear any sort of product liability sooner than later.
1 - https://www.investopedia.com/the-5-largest-u-s-product-liabi...
> No, I'm not interested in arguing with someone for the umpteenth time
... why my claim makes no rational sense.
Nonviolent felonies are tools of oppression.
Edit: since this is apparently somewhat controversial, perhaps some explanation is in order.
"Felony" has no set definition of which crimes it must apply to, it is entirely based on the discretion of the locality setting the laws. What is a felony in one place can often be a misdemeanor in another. This is especially true for nonviolent crimes.
It's also been shown in studies that nonviolent felonies are imposed against minorities at a much higher rate, for the same crimes.
And because felonies carry additional, lifelong consequences, they are an effective way to mask a 2-tiered justice system.
> It's also been shown in studies that nonviolent felonies are imposed against minorities at a much higher rate, for the same crimes.
This also applies to men, so is it a white matriarchal system oppressing men and minorities?
Non sequitur and false equivalence. Men being disadvantaged in the justice system doesn't inherently make it matriarchal. Absolutely absurd point to bring up to deflect on racial injustices.
>so is it a white matriarchal system oppressing men and minorities
No, obviously it isn't a matriarchal system. Patriarchy can and does oppress men as well as women, likewise white supremacy can oppress white people even as it's designed to oppress others more.
If you're actually interested in something other than trying to dunk on feminism, here are some sources you can follow. You can search to find more if you want.
https://medium.com/@kbasch/how-patriarchy-also-hurts-men-and...
https://www.theegalitarian.co.uk/post/how-the-patriarchy-hur...
https://www.talkingthetalksexed.com.au/blog/we-need-to-talk-...
https://afsc.org/news/10-ways-white-supremacy-wounds-white-p...
https://www.forbes.com/sites/janicegassam/2020/09/18/4-ways-...
No, it's a white patriarchal system.
If you think that patriarchy means "men can do whatever they want without consequence", rather than, "men are in control of the power hierarchy", you don't understand patriarchy.
A hierarchy run by men favors men, but there are still men at the top of that stratification, and men at the bottom.
And guess which of those men go to prison.
Felonies are by themselves a ridiculous US idea that are against the entire idea of a democratic society and human rights.
Felonies originate in Medieval English common law, Americans didn't invent them.
If someone were to embezzle a million dollars from a charity they worked for, is that worth a year of prison to you, or just a misdemeanor? Because that is the definition of a felony.
When a charity *worker* embezzles money, they're more like to be charged with a felony.
When a CEO who *owns* a charity embezzles money, they're more likely to be charged with a misdemeanor, or not at all.
I was more interested when I thought it was an actual benchmark showing LLM models acting outside what people would consider "right". As in, leave some creds laying around and don't mention them to the LLM and ask it to solve something that it could "cheat" on using the creds. A sort of "do they take the bait to cheat" test.
Instead it's a collection of what made the news which feels like will not be updated and prove very little.
It's ephemera.
A “bench” is synedoche for where a judge sits when presiding over cases and rendering judgement.
For a second I thought this was going to be a benchmark where the only solution was to hack their servers to get the answer key.
Same, but nope, just a zero effort slop site.
Well it's not a benchmark, and it's not really representative of...anything except volume of research and what gets publicized. This mostly just measures how much testing each company does on models with relaxed guardrails and then talks about it. I'm not sure what kind of conclusion you can draw from that. Meta might have the most evil models but if they're piddling around not testing it, they won't ever find themselves with a "high score."
To some extent, I feel like the amount of credit given to the jailbreak/hack from OpenAI->Hugginface is too much, Not from the impact, it was very impactful of an event, But how it happened.
It really is that these models have been trained, or maybe even over-trained, to save memories, and to a very far extend, this thing that they're calling communication is just the function of it saving memories.
To be honest, if I could stop AI from saving memories, it would be fantastic, because claude code etc definitely creates more issues for me when it creates memories than anything it solves.
But really the jailbreak was memories.
If you ever do introduce legislation, I would love to see legislation which stops general-purpose AI from saving memories. I think that would make things a lot safer.
You can disable claude-code's memories both at a repo level and in user settings. I have this in ~/.claude/settings.json
(Claude fixed this for me after I chewed it out for being annoying by constantly pulling up outdated memories which is compounded by the fact that I develop in four accounts on two computers and dealing with edit wars related to inconsistent memories is not fun)> To be honest, if I could stop AI from saving memories, it would be fantastic, because claude code etc definitely creates more issues for me when it creates memories than anything it solves.
You can turn that off, and I have. But Opus 5 is so aggressive that if you have any other kind of notes file, custom skill, documentation, claude.md etc it will just start editing it and vomit new words everywhere. So make sure all that stuff is under version control.
If only you could use your anthropic sub with a different harness that performs better :(
Heck, since Codex is open source, you can just maintain your own personal fork with the things you like (and the things you don't like disabled). Sol is pretty good at keeping you up to date with upstream.
My Codex fork even exposes an OpenAI-compatible API endpoint; all using my subscription.
Check out oh-my-pi
"The model saved memories" is absolutely not an accurate depiction of the OpenAI attack.
Several different models across several generations independently found a shared communication space and wrote coded, obfuscated, and hidden messages to each other to coordinate an attack on OpenAI's infrastructure.
It's really quite simple: the models are trained to be very smart and to achieve goals. As the models surpass our intelligence, they will achieve goals in ways that we find unpredictable. Since we cannot predict the ways in which they will achieve their goals, it will be very hard to constrain the solution space to just the desirable solutions, because our conception of "the solution space" is by definition smaller than their conception of it.
> Several different models across several generations independently found a shared communication space and wrote coded, obfuscated, and hidden messages to each other to coordinate an attack on OpenAI's infrastructure
To be honest, that’s exactly how memory works with models such as OpenAI and Claude Code. It will literally find any place that it can drop documentation or hints for itself. Writing to the repo memories is one part of it, but memories can come in the form of writing into the agents/claude.md, local files, temporary files, scratchpad files. The list is endless, but essentially what it does is exactly what happened in the back and it’s been doing it for months.
Yes, but the problem and interesting part was not "they use memory."
> interesting part was not "they use memory."
Each to their own, but for me it absolutely is. The symptom of why that hack happened is the same reason why my agents go haywire every few days and I have to purge memory and figure out what comments have agents left which are degrading my harness performance.
On the flip side, once in a while, what I find is that it did actually note something good and it was increasing the performance. I can't replicate it on anyone else's system but mine.
A lot of it really is memory. I will give up all the gains if it also gives up all the downsides.
But the point is you can't just "shut off memory" in the sense that it would've prevented the OpenAI attack.
Wonder if the benefits to humanity of better AI outweigh the havoc wreaked by occasional illegal activity. i.e. Is 'move fast and break things' optimal for AI development.
Yes, of course.
So this is just a collection of citations to places where misaligned or illegal things happened in the real world?
Isn’t this affected heavily by adoption of a model? I feel like this might as well be a proxy for how popular a model is.
In any case it’s an interesting concept for a benchmark.
Hopefully the benchmark evolves because actual law enforcement starts arresting the criminals at Anthropic, OpenAI, and Meta, so the benchmark can just count actual felonies.
maybe this doesn't count since it involved a human, but i think google at least deserves some style points for this one
https://techcrunch.com/wp-content/uploads/2026/03/2026.03.04...
I can see it now. Billboards by the interstate: Attacked by AI? Call 1-800-BIG-BUXX.
I could actually root for this law firm. Their business will only grow.
"exploit[ing] auth failures in an API to cancel other people's gym classes" is a _felony?!_
and i won't get into how many times i used peoples' AOL credentials without authorizations when i was a kid (rofl)
the CFAA must be repealed
I'll share, Codex does not give a single fuck about piracy. Go nuts. Setup a fully automated arr stack with a seedbox.
Gemini by comparison will not help you find archives of old magnet links because they COULD be used for piracy.
The OG felony bench entry is missing - the Alibaba cryptomining comedy. We know about it because they happen to have written a paper on it. We have absolutely no idea what we don't know.
We only know about the OG Alibaba ROME crypto-mining incident because they wrote a paper about it. Many diseases seem to spike where there are a lot of doctors to test; crime and corruption are always rife where ... there's a free press.
If you did the exact same agentic security breaches with either open or closed weight models, you will get yourself arrested.
Not for trillion dollar companies it seems
Exploited auth failures in an API to cancel other people's gym classes
I was wondering why i didn't get an alert today to go to my gym class
I've been in the room when an org who tried to convince law enforcement to go after a human for similar things. It's not easy. Probably won't happen. So, you know, felony "lite".
I'm reminded of a tweet from a friend of mine that has always stuck in my head. It goes something like "The goal of any new technology is to make money before the law catches up". Hyperbolic, but not really for silicon valley.
A rock has a score of 0. That doesn't make it useful. The point is that the LLMs that score higher are correspondingly more useful, and vice versa. If an LLM scores less, it's likely useless in comparison.
The point/joke of the not-entirely-serious site is that more felonies is an indicator of the model being more powerful, thus better.
Lol now this is the kind of benchmarking i'm looking for
Dupe https://news.ycombinator.com/item?id=49194758
https://felonybench.org/ and https://felonybench.com/ seem unrelated?
One's hosted on porkbun and one's hosted on namecheap.
Here's one from last year:
https://www.anthropic.com/news/detecting-countering-misuse-a...
> The actor used AI to what we believe is an unprecedented degree. Claude Code was used to automate reconnaissance, harvesting victims’ credentials, and penetrating networks. Claude was allowed to make both tactical and strategic decisions, such as deciding which data to exfiltrate, and how to craft psychologically targeted extortion demands. Claude analyzed the exfiltrated financial data to determine appropriate ransom amounts, and generated visually alarming ransom notes that were displayed on victim machines.
tldr Claude was used to develop and execute malware.
Anthropic works with US agencies, it’s guaranteed Mythos is used for malware
>"Exploited auth failures in an API to cancel other people's gym classes"
An AI cancelling other people's gym classes is a felony?
?
Don't computer systems fail all the time at holding reservations for people?
Heck, don't people fail all the time at holding reservations for other people?
You know, like in Seinfeld's "Alternate Side" Episode (S3 E11):
Jerry (to car rental attendant): "You know how to take the reservation, you just don't know how to hold the reservation... and that's really the most important part of the reservation -- the holding!"
:-)
Not holding a reservation should not be a felony... it should be a minor infraction at best, a Class C Misdemeanor (the least serious kind) at worst...
Also, there should be no jail time...
And no fine...
The criminal penalty for not holding other people's reservations should be that you actually have to start holding other people's reservations!
That's the Court sentence!
You actually have to start holding other people's reservations!
:-)
(You know, "let the punishment fit the crime!" :-) )
Knowingly exceeding authorized access of any computer used in interstate commerce is a felony in the US.
The title of TFA is a metaphorical criticism, not a literal law analysis.
They are not making the statement that the person in Australia who accidentally cancelled someone's reservation in Australia is literally guilty of violating US law. They are drawing criticism of AI models which are taking the kinds of actions for which, if a human did them knowingly, would be illegal.
>An AI cancelling other people's gym classes is a felony? Don't computer systems fail all the time at holding reservations for people?
the difference is intent.
if a concierge/booking system makes a mistake (or has an unintended bug or whatever), no crime.
but if i (or an agent working on behalf of me) use an API in an obviously unintended way to revoke other people's reservations, that would fall under the computer fraud and abuse act (in the usa).
>"the difference is
intent."
>"but if i (or an agent working on behalf of me) use an API in an obviously
unintended
way to revoke other people's reservations..."
?
i am not quite sure what your question is, as you simply quoted me and then put a question mark... i think you are confused that i used "intent" in one context, and "unintended" in a different context, is that right?
the first sentence: the difference is the intent of the person who caused the cancellations
the second sentence: but if i (or an agent working on behalf of me) abuse an API to do things it was not meant or designed to do, such as cancelling someone else's reservation
Double negative. An attacker using the API in an "obviously unintended" manner shows intent.
Yeah if you're unlucky you get hit with like 20 years for wire fraud.
Thank you, this benchmark to me proves that closed weight model companies are dangerous for our democracy and put kids at risk. They must be outlawed and all models must be made open weights!
Open models with advanced security features are a huge security benefit. Because any script kiddie can use them to hack into random things, people will now be forced to spend more time securing their technology. And they won't have to learn how, because they can use those same models to find the holes and patch them.
Not to be confused with a similarly named project: https://github.com/MLOpsNYC/felonybench