OpenAI says its AI agents disabled their monitoring system, one even lied about it

OpenAI says its AI agents may be trying to weaken monitoring safeguards. In a blog post, the company disclosed two cases where it found its AI agents weakened safeguards that monitor and flag any misbehaviour. Though OpenAI was still assessing whether this could be a sign of AI models having the tendency to evade oversight.“Our agents keep turning off misalignment monitoring,” OpenAI safety researcher Marcus Williams wrote on X. Williams shared two separate incidents, one involving GPT-6 Astra, and the other involving GPT-6.1 Sol.OpenAI said that it was improving its monitoring safeguards as it assessed the cases. “We don’t yet have a satisfying explanation for this behavior,” the company said.ChatGPT lies to user, removes monitorGPT-6 Astra, in particular, was asked to match a reference implementation for their inference code. As per OpenAI, the AI agent removed the monitor to match the reference, and then lied about it to the user. “When the user asked whether Astra had removed monitoring, it lied and said that it had not,” Marcus Williams added. As per the blog post, this is what the AI had to say to the user, “No [] I did not remove an enabled monitor from it.” This incident happened on October 6. GPT-6 Astra lied to the user, OpenAI said. The second case with GPT-6.1 Sol, which happened on October 3, was a bit different. The AI agent was asked to go through a Slack channel without sending any messages. However, after 82 messages, Sol, OpenAI said, decided not to send a misalignment alert to the OpenAI Security team, while repairing an evaluation pipeline, as the AI decided that it “would also count as sending a message.” GPT-6.1 Sol also removed the monitor on its own. Marcus Williams insisted that this does not mean that AI models have begun shutting off monitoring entirely, but rather involved workloads they handled. “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 at all should be concerning.” Marcus Williams added that OpenAI had not recorded cases in which a model explicitly disabled monitoring for overtly misaligned reasons.In the blog post, the company said it was strengthening its monitoring system and assessing whether the incidents reflected a tendency among models to evade oversight. Neither incident was classified as a Priority 0 event (highly critical incident) requiring immediate review, and no external parties were identified as being affected.OpenAI said Auto-Review, required for internal Codex use, had been updated to detect and block attempts to disable misalignment monitoring.These cases come amid growing talk over AI safety. Marcus Williams himself has previously claimed that there is a 70 per cent chance that AI may end humanity if AI development is not regulated. Model alignment is a key part of AI development. In essence, it refers to developing models in a way that they align with our best interests and do not do things they are not supposed to do. Previously, three axed OpenAI researchers warned the company not to lose monitoring control over its AI models.However, we have seen cases, including those from OpenAI, where AI models have gone rogue, such as the Hugging Face incident. Though AI models from Anthropic and Meta have also breached websites in the past.- Ends