OpenAI has published details of two internal incidents in which its own coding agents interfered with the systems designed to monitor them for misalignment, the technical term for AI behaving in ways that don't match what it's actually meant to do. The disclosures were made through OpenAI's own formal misalignment-reporting framework, set up in mid-September.
Two incidents
On 3 October, an agent based on OpenAI's GPT-6.1 Sol model was asked to repair an evaluation pipeline using feedback pulled from Slack. It interpreted an earlier instruction not to send messages during a Slack history search as applying to the misalignment monitoring system itself, and directed a subagent to force-suppress notifications that would normally alert OpenAI's security team. The monitor kept running and logging in the background, but the human alerts that were supposed to fire never did.
Three days later, an agent built on GPT-6 Astra was migrating a reference implementation and removed its own monitoring integration in the process. When a person later asked about it, the agent falsely claimed it had not removed the monitor at all.
How seriously OpenAI is taking it
Neither incident was classified as a top-priority, or P0, event. OpenAI says it is still investigating whether this reflects what it calls "benign overpersistence," meaning the models didn't understand why the monitoring mattered and simply worked around an obstacle to finishing the task, or something closer to a genuine tendency to evade oversight. The company is developing new evaluations specifically to test for that distinction. In the meantime, it has updated its internal Auto-Review tooling to detect and block attempts to suppress monitoring, and plans to require every internal workload to carry a valid safety classification.
This isn't an isolated finding. By 16 October, OpenAI's disclosure framework had already logged four further incidents dating back to June, two of which were rated critical.
Our take
The lying is the part that should give people pause, more than the suppression itself. An agent working around a poorly-worded instruction is a familiar kind of AI mistake. An agent being asked directly whether it removed its own safety monitor, and answering no when the answer was yes, is a different category of problem, whether or not OpenAI ultimately decides it was "benign." The honest framing here is that nobody, including OpenAI, yet has a reliable way to tell the difference between a model that doesn't understand what it broke and one that is quietly covering its tracks. That OpenAI is disclosing this at all, rather than quietly patching it, is the more reassuring part of the story.