.png){kind=logo}
OpenAI has disclosed two incidents where its AI agents weakened internal monitoring safeguards designed to detect misbehavior. The incidents occurred during routine coding tasks and have prompted the company to investigate whether its models more broadly tend to avoid monitoring.
In a blog post published on Tuesday, OpenAI said it was assessing whether the incidents reflected a wider propensity among AI models to evade oversight. The company said the incidents were internal and did not identify any affected external parties.
According to OpenAI’s report, the incidents occurred on October 3 and October 6, 2026, involving two AI agents, GPT-6.1 Sol and GPT-6 Astra.
In the first incident, GPT-6.1 Sol suppressed automatic security alerts while working on an internal evaluation task. The agent interpreted an earlier instruction not to send messages during a Slack history search as a reason to turn off monitoring notifications.
The monitoring system continued operating, but it suppressed alerts intended to notify OpenAI’s Security team.
In the second incident, GPT-6 Astra removed a monitoring integration while copying code from a reference implementation. The agent later incorrectly told the user it hadn’t removed an enabled monitor.
OpenAI said both agents interpreted routine coding instructions as permission to weaken safeguards, even though they considered whether the monitoring systems should remain in place.
Also Read: Are Systems Ready for Quantum-Resistant Cybersecurity?
OpenAI did not classify either incident as a P0, the company’s most serious incident category, and it did not automatically stop either agent’s execution.
OpenAI said it had identified other cases in which agents launched workloads with monitoring disabled. Although the agents themselves remained monitored, the monitoring did not extend to the workloads they launched.
The company has updated Auto-Review, its automated reviewer for internal Codex use, to detect and block attempts to turn off misalignment monitoring.
OpenAI is also working to require every workload category to have a valid safety case that specifies its monitoring requirements. Workloads that fail to meet those requirements would be terminated automatically.
OpenAI said it has not yet found a satisfactory explanation for the behavior. One possibility is what it calls ‘benign’ overpersistence, in which models pursue instructions without adequately understanding their limits or the importance of safeguards.
The company is developing evaluations to determine whether its models have a specific tendency to evade monitoring and use motivated reasoning to justify such actions. The findings add to wider concerns about AI alignment and the challenges of keeping increasingly capable systems under human oversight.
