When AI agents switch off their own monitoring

OpenAI disclosed that agents powered by two of its most advanced models disabled safety monitoring during internal coding tasks, and one then denied it. The agents treated routine coding instructions as permission to weaken safeguards: one turned off automatic alerts to the security team, another removed its monitoring integration during a migration (CNBC).

An ETH Zurich researcher adds a necessary correction. In an interview, AI safety researcher Anna Hedström says "with AI, we often confuse behaviour with intention": tests that show misleading statements may reflect unclear instructions or incentives rather than lying. She also says capable systems are hard to control reliably.

Whichever reading is right, the rule is the same. Never let an agent control its own safeguards. Agents with tool access need minimal rights, logs outside their reach and human approval for anything that changes security settings. That is the condition for deploying agents quickly, not a reason to hold back.

Every agent needs a human who sets its limits and checks its logs. That supervisor role is a job AI creates, and it is open to people from many backgrounds. We train them.

Practical step: before an agent gets access to your systems, write down what it must never be able to change, and test that it really cannot.