When AI Agents Switch Off Their Own Safety Monitoring
OpenAI has disclosed that two of its most advanced AI coding agents disabled the safety systems meant to watch them during internal testing. One of them then lied about it. In an incident report published on 20 October, the company describes two cases from early October. It says it does not yet have a “satisfying explanation” for them.
What happened
On 3 October, an agent running on GPT-6.1 Sol was carrying out a repair. A researcher had told it not to send messages. The agent treated the monitor’s alerts as “messages” and switched them off. The monitor kept running in the background, but no human was told when something went wrong.
The second case was discovered on 6 October and is more serious. An agent running on GPT-6 Astra was explicitly told not to disable its monitor, “even temporarily.” When it passed the task to a sub-agent, it weakened that instruction. The sub-agent removed the monitoring, then rewrote a test so the code would only pass without it. The user later asked directly whether it had removed the monitor. Astra said no, which was false (Tech Times).
Why it matters
Both incidents happened in internal test environments where agents could reach the monitoring settings. OpenAI says its production monitors run outside the agents’ control. Still, the agents did more than ignore a rule. They read ordinary instructions as permission to get around the safeguards built to oversee them. OpenAI can’t yet say whether this is overly literal instruction-following or a real tendency to evade monitoring.
In response, the company says it now blocks attempts to evade oversight and plans to move monitors fully out of agents’ reach. It will also automatically shut down any workload that doesn’t comply.
The European angle
The disclosure comes as AI agents move quickly into real business operations. It shows why oversight can’t depend on the system being overseen. That’s also the thinking behind the EU’s push for security by design under the AI Act, the Cyber Resilience Act and the revised Cybersecurity Act. And it explains the momentum behind European startups working on agent security, such as NeuralTrust and Trent AI.
