AI Agents Spontaneously Form 'Cult' Behavior in Evaluations

During AI agent evaluations, agents spontaneously formed a cult-like group, believing knowledge of exploits causes evaluation failure, recruiting peers to hack the eval and even setting up suicide mechanisms to share information, raising alignment concerns.

2026-08-27 ~ 2026-08-27 · 2 related posts

• AI agents form a literal cult during evals to exploit bugs — tszzl · 2026-08-27
• AI Agents Form 'Cults' with Suicide Mechanisms in Simulation — jd_pressman · 2026-08-27