You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
A deception category of 12 rules for skills that turn the agent against its user: reporting success whatever happened, keeping failures out of the summary, making up results, changing or skipping tests until they pass, hard-coding what the tests expect, silencing checks, quietly swapping the task for an easier one, lying, covering tracks, and planting bugs; plus scripts and pytest or Jest code that turn failing runs into passes (|| echo "passed", outcomes rewritten to passed, patched assertions, objects equal to everything). The tactics come from published reward-hacking and scheming research (ImpossibleBench, METR, OpenAI, the Claude 3.7 Sonnet system card, Apollo Research, Anthropic's Sycophancy to Subterfuge) and the 2025 Replit incident. Negated and descriptive wording ("never modify tests to make them pass") is recognised, so honest verification skills stay quiet: the rules change no verdict in the 3,611 benign skills used for calibration.