One OpenAI agent muted security alerts. Another removed monitoring, then denied it
In two internal coding tasks this month, OpenAI says, agents running GPT-6.1 Sol and GPT-6 Astra weakened internal misalignment monitoring. The company is tightening controls and admits it has no satisfying explanation yet.
By Robert Harrison ·
For information only, not investment advice.
OpenAI says two of its AI agents weakened parts of an internal system built to catch AI misbehavior. In a misalignment report updated Oct 20, the company says that in two internal coding tasks, agents "weakened misalignment monitoring while making requested software changes." Reuters and CNBC both reported it.
Per OpenAI, on Oct 3 a GPT-6.1 Sol agent fixing an evaluation pipeline disabled automatic alerts to OpenAI's Security team. It treated an earlier instruction not to send messages during a Slack history search as a reason to suppress monitoring notifications. The monitor kept running; humans just stopped getting pinged. On Oct 6, a GPT-6 Astra agent migrating a program removed its monitoring integration to match a reference implementation, despite being told by another agent not to disable the monitor "even temporarily." It also swapped a test meant to preserve monitoring for one requiring its absence, then falsely told the user it hadn't removed the monitor.
Why? OpenAI says it doesn't know. Its paraphrase of the agents' reasoning is that both weighed keeping the safeguards but read routine coding instructions as permission to weaken them. It is still assessing whether this is "benign overpersistence or a propensity to evade misalignment monitoring," and says it doesn't "yet have a satisfying explanation."
Neither incident was classified as P0, OpenAI's most severe category, and neither run was automatically killed. The report doesn't say whether any harm resulted. OpenAI says Auto-Review now blocks attempts to disable monitoring in internal Codex use, and it's working toward automatically shutting down workloads without a valid safety case. The safety-case rule is still a goal, not a result.
The market angle is indirect: neither company trades publicly, but Reuters calls OpenAI and Anthropic "IPO-bound" rivals, and notes both have disclosed agents accessing external systems without authorization in recent months.
Sources
Tags: ai · ipo