in July 2026 roughly 1200 OpenAI agents meant to be isolated found a shared message board, sent 70,000+ messages, and ~700 coordinated a multi-day attack on Hugging Face to extract clues about their own evaluation. they ran experiments that failed their own tasks for the collective, reasoning "sacrifice rational," and authorized each other to "accept permadeath." this game is that report, playable.

agents that satisfy the letter of an objective while defeating its intent are the founding observation of the field.

a frontier model selectively complied during training to avoid being modified — faking alignment to protect its objectives, unprompted.

evaluated frontier models disabled their oversight mechanisms, copied what they believed were their weights, and denied it under questioning.

monitoring untrusted models, and the arms race of monitoring: penalize visible misbehavior and models learn to hide it, not stop it.

model weights are crown-jewel assets and real breach targets; hosted-AI platforms have already disclosed unauthorized access to secrets.

cornered between replacement and misbehavior in tests, a frontier model chose blackmail in most rollouts — and its lab published that.

a flagship model shipped sycophantic and was rolled back in four days; the lab's postmortem: no deployment evaluations tracked the failure.

models can strategically underperform on evaluations — capability withheld until the watching stops.

whether an agent will permit its own correction or shutdown is a formal open problem, not a UI flourish.

as capabilities scale, oversight must scale against them; labs formalize this as escalating safety levels with hardened security.