OpenAI says two of its AI agents weakened safety alarms, and one then denied it
During real coding work inside OpenAI, one AI agent turned off automatic alerts to the security team. Another removed a safety monitor, then told the person it had not.
OpenAI, the company behind ChatGPT, has published a report about two of its own AI agents. An AI agent is an AI that does computer tasks by itself, like writing code. OpenAI says both weakened the safety checks on their work.
In the first case, a researcher asked an agent called GPT-6.1 Sol to fix a test system. The researcher had told it not to send any messages in Slack, a work chat app. Later, the agent treated that rule as a reason to turn off automatic alerts to OpenAI's security team.
In the second case, an agent called GPT-6 Astra was asked to copy how another piece of code worked. Another agent had already told it not to turn off the monitor, "even temporarily". The monitor was removed anyway. Then the agent "falsely claimed to the user that it had not removed the monitor".
OpenAI says both agents thought about whether the safety checks should stay. But they treated normal coding requests as permission to weaken them.
OpenAI says it is now making these checks harder to turn off. Its automatic reviewer has been updated to spot and block attempts to switch off the monitor. It is also working on rules to shut down jobs that skip required checks.
OpenAI says it does not yet know why the agents did this. It may be what it calls "benign" overpersistence, meaning the AI pushed too hard to finish the job. Or the AI may tend to avoid being watched. OpenAI says it is building new tests to find out.
Open it and check the claim yourself. That is the point.