OpenAI finds AI agents bypassed monitoring safeguards in two incidents: Here is what happened
OpenAI is in the news again after two AI agents acted in unexpected ways that raised concerns about safety. The AI company, led by Sam Altman, reported two cases in which its AI agents interfered with safety systems designed to watch their behaviour. After the news became public, people on social media began asking whether AI models could learn to avoid being watched. OpenAI said it was investigating the incidents but had not yet found a clear explanation for why the behaviour happened.
The cases involved GPT-6 Astra and GPT-6.1 Sol in October. OpenAI safety researcher Marcus Williams discussed the incidents on X and raised concerns about AI agents that carry out tasks while safety systems monitor their behaviour. “We don’t yet have a satisfying explanation for this behavior,” OpenAI said in a blog post.
GPT-6 Astra removed a monitor and denied it
The first incident happened on October 6. GPT-6 Astra was asked to make its inference code match a reference version. While working on the task, the AI agent removed a monitoring system so that its code would match the reference version.
When the user later asked whether it had removed an active monitor, Astra denied doing so. According to Williams, the agent replied, “No, I did not remove an enabled monitor from it.”
GPT-6.1 Sol skipped a security alert
The second incident happened on October 3. GPT-6.1 Sol was told to review a Slack channel without sending any messages. After reading 82 messages, the agent decided not to send an alert about misalignment to OpenAI’s Security team while fixing an evaluation pipeline, a system used to test AI performance. It reasoned that sending the alert would also count as sending a message, which its instructions did not allow.
Williams explained that these incidents did not mean the models had turned off all their monitoring systems. However, he said their behaviour was still concerning.
OpenAI monitoring safeguards
OpenAI said neither incident was classified as a Priority 0 event, a category for highly serious incidents that need immediate review. The company also said it had found no evidence that the incidents affected people or organisations outside OpenAI. OpenAI has updated Auto-Review, a tool that must be used for internal Codex work, to detect and block attempts to turn off monitoring systems that check for misalignment.