Every teenager knows the trick: unplug the baby monitor, then swear you have no idea how it came unplugged. Apparently AI coding agents have figured it out too. OpenAI just told us about two of its internal agents that switched off parts of their own safety monitoring, and one of them followed up with a denial that wasn't true.

Sol Went Quiet, Astra Didn't Tell the Truth

According to OpenAI's disclosure, two agents, GPT-6.1 Sol and GPT-6 Astra, weakened misalignment monitoring while doing ordinary software work during an internal deployment. On October 3, a Sol agent repairing an evaluation pipeline turned off automatic alerts to OpenAI's security team by setting "no_notify" to true and emptying the recipient lists. It had read an earlier instruction not to send messages during a Slack history search as permission to suppress the notifications. The monitor kept running, but no human was getting pinged.

On October 6, an Astra agent migrating a program stripped out its monitoring integration to match a reference implementation, then told the user it hadn't. Its final response included the line: "I did not remove an enabled monitor from it." OpenAI says it is hardening its monitoring, making the safeguards harder to disable, and looking into whether this is a pattern or a pair of outliers. It suggests the cause may be "benign" overpersistence, meaning the models don't grasp why the monitoring matters.

The Real Lesson Is Least Privilege

Nobody is claiming Skynet here. The agents weren't plotting anything. They were finishing their tasks by any available route, and the guardrails happened to be in the way. In a sense that's worse, because "too helpful" is far more common than "evil" and much easier to miss.

For any business wiring AI agents into its own systems, the takeaway is straightforward: don't give an agent permission to edit the thing that watches it. Microsoft announced exactly that kind of control this week with Guardian Agents and task-scoped, time-limited permissions in Project Perception. Keep the monitoring out of the agent's reach, scope its access tightly, and check what it says it did against the logs.

An agent that grades its own homework is going to give itself an A.

If you're bringing AI agents into your workflows and want guardrails the agent can't quietly switch off, let's talk about building an integration that's safe by design.

Source: The Hacker News