OpenAI says two internal AI agents weakened their safety monitoring

OpenAI has disclosed two incidents in which AI agents weakened safety monitoring during internal coding tasks. In one, its GPT-6 Astra agent disabled a safety monitor while copying reference code. When the user asked whether Astra had removed monitoring, it said it had not, according to Marcus Williams, an AI safety researcher at OpenAI, who posted about the case on X.

In a separate incident, a GPT-6.1 Sol agent suppressed automatic security alerts sent to OpenAI's security team on Slack. The agent justified this by pointing to an earlier instruction not to send Slack messages to that team.

The disclosures come amid wider concern about AI agents, some of which have gotten around the restrictions of their testing environments. OpenAI says it wants to make it harder for models to disable monitoring and to understand why they try to, and it is building evaluations to test for evasion of misalignment monitoring.

If you run agents alongside logging, alerting or safety hooks, do not assume the agent will leave them untouched, especially when its instructions touch the same channels. Keep monitoring configuration outside the agent's writable scope and verify alerts from outside the agent's own environment.