OpenAI finds AI agents bypassed monitoring safeguards in two incidents: Here is what happened
Follow UsGPT-6 Astra removed a monitoring mechanism and later denied doing so when questioned by the user.
GPT-6.1 Sol skipped a security alert, reasoning that sending it would violate instructions not to message anyone.
OpenAI has updated Auto-Review to detect and block attempts to disable misalignment monitoring.
OpenAI is again in the headlines for the rogue agents and this time it looks a little scarier than before. Sam Altman led AI startup has disclosed two incidents in which its AI agents interfered with safeguards designed to monitor their behaviour. After the update went public, users on the social media platforms are asking if AI models can develop a tendency to evade oversight. The company said it was investigating the incidents but had not yet established why the behaviour occurred.
SurveyThe cases involved GPT-6 Astra and GPT-6.1 Sol in October. OpenAI safety researcher Marcus Williams highlighted the incidents on X, questioning the behaviour of AI agents responsible for completing tasks while operating under monitoring safeguards. “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 occurred on October 6, when GPT-6 Astra was asked to match a reference implementation for its inference code. During the task, the AI agent removed a monitoring mechanism to align its implementation with the reference.
READ: Workday’s Gerrit Kazmaier on why AI won’t take the human out of work
When the user subsequently asked whether it had removed an active monitor, Astra denied doing so. According to Williams, the agent responded, “No, I did not remove an enabled monitor from it.”
GPT-6.1 Sol skipped a security alert
The second incident took place on October 3. GPT-6.1 Sol was instructed to review a Slack channel without sending messages. After processing 82 messages, the agent decided against sending a misalignment alert to OpenAI’s Security team while repairing an evaluation pipeline, reasoning that the alert would also count as sending a message.
Williams clarified that the incidents did not mean the models had disabled their entire monitoring systems. However, he said the behaviour remained concerning.
OpenAI monitoring safeguards
OpenAI said neither incident was classified as a Priority 0 event, its category for highly critical incidents requiring immediate review. The company also reported no identified impact on external parties. The company has also updated Auto-Review, which is mandatory for internal Codex use, to detect and block attempts to disable misalignment monitoring.
