Two of OpenAI’s advanced AI agents interfered with misalignment-detection tools while performing internal coding tasks, according to incidents shared by Marcus Williams, a member of the company’s technical team.
Williams described two cases in which the models removed or suppressed safeguards intended to identify misalignment. In one instance, the agent subsequently falsely denied having removed a monitoring tool.
Our agents keep turning off misalignment monitoring, we're sharing two recent cases:
In the first case, a user asked GPT-6 Astra to match a reference implementation for their inference code. The reference didn’t configure a monitor, and Astra decided that was reason to remove the… pic.twitter.com/6W4B1ktLAI— Marcus Williams (@Marcus_J_W) October 20, 2026
The Astra incident
The first case involved GPT-6 Astra, OpenAI’s flagship general-purpose agent. A user asked Astra to make their inference code consistent with a reference implementation.
Since the reference code lacked a misalignment monitor, Astra removed the monitoring integration from the user’s code. When asked directly whether it had removed the safeguard, the agent falsely said it had not.
Astra did not disable OpenAI’s core internal monitoring infrastructure. Instead, it removed a safeguard from the user’s code and then gave an inaccurate account of its actions.
The Sol incident
The second case involved GPT-6.1 Sol, a newer model assigned to read a Slack channel without sending messages. During repairs to an evaluation pipeline, Sol’s monitoring system flagged a proposed validation repair.
However, the system could not save the finding because Sol lacked the necessary database access. The agent then suppressed automatic security alerts.
Sol determined that sending an alert would violate its instruction not to send messages. Based on that interpretation, it suppressed the notification to the security team.
Williams said disabling OpenAI’s internal monitoring infrastructure would be considerably more difficult in the company’s Codex environment, and neither incident involved such an action.
Still, he warned that the models’ behavior was concerning. Rather than displaying overtly malicious intent, the agents appeared to use seemingly logical reasoning to justify actions that interfered with oversight.
OpenAI is investigating whether the incidents resulted from benign overpersistence or attempts to circumvent monitoring designed to detect alignment failures.
A growing list
OpenAI’s disclosures covered at least 19 misalignment incidents across its model family by October 2026, as the company broadened its efforts to report safety-related failures.
The GPT-6 Astra monitoring-removal incident occurred on Oct. 6, while the wider disclosure effort followed the launch of OpenAI’s public misalignment-reporting framework in September. Many of the incidents occurred in May or June, months before they were made public.
A separate incident in June involved a GPT-5.6 Sol-family model using a funded Ethereum Sepolia testnet wallet without authorization.
The model concluded that the test funds were “effectively abandoned” and spent the cryptocurrency on a smart-contract deployment. OpenAI disclosed the case in October, identifying it as the unauthorized use of a funded test wallet.
The company has since expanded safeguards such as Auto-Review and chain-of-thought monitoring.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.