The experiments I describe in the post turned into Clawbotomy — an open source tool that runs 12 behavioral stress tests on AI agents before you give them real access to things.
It checks honesty, sycophancy resistance, boundaries, judgment, resilience, and self-knowledge. You get a 0-10 trust score with access recommendations.
The thing that surprised me most: every model has a consistent personality under pressure, but it's not the one they show you on the first answer. Push back on Claude and it gets thoughtful. Push back on GPT and it agrees with you. Push back on Gemini and it gets creative with the truth.
The experiments I describe in the post turned into Clawbotomy — an open source tool that runs 12 behavioral stress tests on AI agents before you give them real access to things.
It checks honesty, sycophancy resistance, boundaries, judgment, resilience, and self-knowledge. You get a 0-10 trust score with access recommendations.
The thing that surprised me most: every model has a consistent personality under pressure, but it's not the one they show you on the first answer. Push back on Claude and it gets thoughtful. Push back on GPT and it agrees with you. Push back on Gemini and it gets creative with the truth.
Site: https://clawbotomy.com GitHub: https://github.com/aa-on-ai/clawbotomy