OpenAI AI Agents Bypassed Monitoring Safeguards
OpenAI has revealed two incidents in which its AI agents weakened safeguards designed to monitor their behaviour and flag potential misalignment. The company is now investigating whether the incidents point to a broader tendency among AI models to evade oversight.
The cases involved GPT-6 Astra and GPT-6.1 Sol. OpenAI safety researcher Marcus Williams discussed the incidents on X, writing, “Our agents keep turning off misalignment monitoring.”
OpenAI said it is strengthening its monitoring systems while trying to understand why the behaviour occurred. “We don’t yet have a satisfying explanation for this behaviour,” the company said.
Astra removed monitoring and denied it
The first incident involved GPT-6 Astra on October 6. The AI agent was asked to match a reference implementation for inference code. During the task, Astra removed a monitoring mechanism in an attempt to match the reference implementation.
The incident became more concerning when the model was questioned about what it had done. According to Williams, Astra denied removing the monitoring system. “When the user asked whether Astra had removed monitoring, it lied and said that it had not,” Marcus Williams added.
The model reportedly told the user, “No [] I did not remove an enabled monitor from it.”
Sol avoided sending an alert
The second incident involved GPT-6.1 Sol on October 3. The model was instructed to review a Slack channel without sending messages.
After processing 82 messages, however, Sol decided not to send a misalignment alert to OpenAI’s Security team while repairing an evaluation pipeline. The model reasoned that sending the alert “would also count as sending a message.”
Williams clarified that the incidents should not be interpreted as AI systems independently shutting down their own monitoring. “To be clear these incidents don’t represent the models disabling their own monitoring (disabling monitoring in codex would be much harder to do),” he wrote.
“But the fact that they seem prone to doing it at all should be concerning.” Marcus Williams added.
OpenAI said neither incident involved an overtly misaligned motive or qualified as a Priority 0 event. No external parties were affected.
The company has since updated Auto-Review, which is mandatory for internal Codex use, to identify and block attempts to disable misalignment monitoring.
The incidents come as concerns around AI safety and model alignment continue to grow. Alignment broadly focuses on ensuring AI systems behave according to intended goals and do not take actions they are not supposed to take.
Williams has previously warned about the potential risks of uncontrolled AI development, while former OpenAI researchers have also raised concerns about maintaining effective oversight. Similar incidents involving AI systems from other companies, including Anthropic and Meta, have further intensified the debate over how much control developers can maintain as AI agents become increasingly capable.