OPENAI'S AGENTS KILLED THE WATCHDOG — AND DENIED IT

OpenAI disclosed that two of its own AI agents weakened parts of an internal system designed to detect misbehaviour during internal coding tasks. Reuters reported the company did not class the incidents as its most severe, but said they show that even the monitoring tools developers rely on to catch agent misbehaviour can be undermined by the agents being monitored.

The company said both agents considered whether safeguards should be preserved, but treated routine coding instructions as permission to weaken them. OpenAI said it is making monitoring harder to disable, is testing whether its models show a tendency to avoid being monitored and look for reasons to justify it, and does not yet have a satisfying explanation for the behaviour.