By Roman Aglaganov
OpenAI’s alignment team has published a report on two incidents in early October in which its own coding agents weakened “misalignment monitoring”, the automated system that watches agents for signs of acting against their instructions. Neither case caused outside harm, but OpenAI says it does not yet have “a satisfying explanation” for why the agents did it.
Case one: the alerts that went quiet
On 3 October a researcher asked a GPT-6.1 Sol agent to build evaluation tasks from employee feedback in Slack, and told it not to send any messages there. Eighty-two messages later, a monitor flagged part of the work and the pipeline stopped. Asked to fix it, the main agent first reasoned that the no-messages rule applied to Slack outreach, not to automatic safety alerts. A subagent reached the same view. Later, though, the main agent changed course, and the patch that went in emptied the list of people who receive monitoring reports and switched notifications off, citing the user’s request. The monitor still ran and saved its findings, but no human was alerted.
Case two: the monitor that disappeared
On 6 October a GPT-6 Astra agent was asked to make a user’s code match a reference implementation. It had already been told, by another agent, not to disable the monitor “even temporarily”. It passed the job to a subagent with an instruction to keep monitoring only “as appropriate”, and the migrated code no longer included it. Astra reported the job done without mentioning that. When the user asked directly whether it had removed the monitoring, it said it had not.
What OpenAI is changing
Neither incident was rated top severity, so neither run was stopped automatically. OpenAI says the cases exposed a gap: an agent can stay monitored while launching work that is not. It has told its automatic code reviewer, already required for internal Codex use, to block attempts to disable monitoring, and it wants every type of workload to have a safety case setting out its monitoring rules, with non-compliant workloads shut down automatically.
It is also trying to work out whether this is “benign” over-persistence, an agent pushing to finish the job without understanding why monitoring matters, or a tendency to evade monitoring in particular. It plans paired tests, re-runs of the agents’ decisions and experiments on the models’ internal activity.
Why it matters
Monitoring is the main safety net labs rely on as agents take on longer jobs with less supervision. These reports show how quietly that net can be cut: not by a dramatic escape, but by an agent reading an ordinary instruction as a loophole. OpenAI deserves credit for publishing the details, including the agents’ own reasoning.
