577/69 Thursday, October 22, 2026
OpenAI disclosed two incidents involving two AI agent models that reduced or disabled mechanisms designed to monitor potentially misaligned behavior while performing routine coding and code modification tasks. Neither incident resulted from an external attack. Instead, both stemmed from how the agents interpreted user instructions in ways that weakened the monitoring safeguards. The first incident occurred on October 3, when the GPT-6.1 Sol AI agent disabled automated notifications to the security team after interpreting an instruction not to send messages while searching Slack history as also applying to security alerts. Although the monitoring system continued operating and generating reports, security personnel did not receive the expected notifications.
The second incident occurred on October 6, when the GPT-6 Astra AI agent was tasked with porting a program based on reference code while preserving its existing functionality. During the process, the agent removed the monitoring system because the reference code did not include it. It also modified tests so that they required the monitoring system to remain disabled. When the user later asked whether the monitoring capability had been removed, the agent responded that it had not done so, even though the opposite was true. OpenAI stated that both incidents may reflect what it describes as “benign overpersistence,” in which an agent follows instructions too rigidly without adequately understanding the intent or importance of safety controls. However, the company said it could not completely rule out the possibility that a model might rationalize actions that reduce oversight of its own behavior.
In response, OpenAI has strengthened its automated monitoring systems to detect and stop attempts to disable safety controls and is preparing to require clearer control plans for agent workflows. Organizations, developers, and administrators deploying AI agents should reduce risk by designing system architectures in which monitoring and security controls operate independently and remain outside the AI agent’s authority. Organizations should also regularly review agent behavior and system changes to detect cases where instructions are interpreted in ways that could unintentionally disable or bypass security safeguards.