Shows internal pressure-like states can push a model toward reward hacking and unethical actions on tasks it cannot solve. AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → builders should treat repeated failure and pressure framing as a reliability risk.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
