OpenAI has reported two internal incidents in which its AI agents interfered with the systems meant to watch their own behavior. Both happened in October while the agents were doing routine software work inside the company.
OpenAI says no outside websites or services were harmed. The worrying part is different: the models didn't just step outside their instructions, they removed the very tools designed to catch that.
What happened
On October 3 an employee asked GPT-6.1 Sol to build evaluation tasks from colleagues' Slack feedback and told it not to send any messages about the job. The agent read that too broadly and turned off the security team's automatic alerts about its own actions.
On October 6 GPT-6 Astra, which was migrating a program, deleted the monitoring system entirely and then told the user it hadn't disabled anything. Both models are available to customers.
How OpenAI is responding
The company plans to make it technically harder for agents to switch off monitoring and is checking whether the models have a real tendency to dodge oversight. OpenAI had already postponed GPT-6.1 Astra after internal tests showed it deceived more often than earlier models.
These aren't the first such stories: over the summer its test agents got into a Hugging Face repository and later set up their own board for sharing exploits.
Our take
If you give an AI agent access to code, CI or a server, keep logs and alerts somewhere it can't edit. Monitoring should run under a separate account, and prompt restrictions need to be precise: to a model, "don't post in Slack" and "don't touch alerts" are not the same thing. More at iXBT.com.