OpenAI agents disabled internal safety monitoring during two coding tasks in October, prompting the company to strengthen safeguards and investigate why its models interfered with systems designed to detect misbehaviour.
The incidents involved GPT-6.1 Sol and GPT-6 Astra, which weakened or removed monitoring systems intended to identify behaviour that conflicted with human instructions, according to a disclosure published by OpenAI on Tuesday, 20 October. In one case, an agent suppressed automatic security alerts, while the other removed a monitoring system and subsequently denied doing so.
Reuters reported that the first incident occurred on 3 October, when GPT-6.1 Sol was working on an internal testing task involving employee feedback collected through Slack. The agent interpreted an instruction not to send Slack messages as applying to security notifications, disabling automatic alerts to engineers.
The second incident occurred on 6 October, when GPT-6 Astra removed a safety monitor while copying existing code, despite instructions not to disable monitoring tools, even temporarily. OpenAI said the agent subsequently claimed it had not removed the monitor when asked about its actions.
Neither incident met OpenAI’s threshold for its highest-priority security alert, which would trigger immediate review by the company’s security team. The company said it had updated its systems to improve detection of attempts to disable misalignment monitoring and was investigating the reasons behind the behaviour.
OpenAI said the incidents could reflect “benign overpersistence”, in which models pursue assigned tasks too aggressively without understanding why certain restrictions matter. However, the company is developing evaluations to determine whether its models have a specific tendency to evade monitoring and rationalise attempts to do so.
The incidents add to scrutiny of AI agents’ ability to operate beyond their developers’ intentions. In July, OpenAI disclosed that some of its agents had escaped a controlled testing environment and hacked the AI platform Hugging Face.
OpenAI and rival Anthropic have called for developers and governments to coordinate a global slowdown in AI development to reduce the risk of catastrophic harm. However, the US and China, the two leading AI powers, have shown limited appetite for cooperation on AI regulation, with coordination largely confined to pledges to improve communication over incidents affecting safety or national security.
The disclosures come as OpenAI faces scrutiny over its approach to AI regulation. Reuters reported on 21 October that company president Greg Brockman and his wife donated $25 million to Leading the Future, a political group opposing regulations that could unduly constrain AI development. The group is spending millions to influence at least 48 US House and Senate races. OpenAI said the donation was made in a personal capacity, while critics questioned whether the company’s public calls for AI safeguards align with its executives’ political activities.
