OpenAI's misalignment report (posted Oct 20) covers two internal incidents. On Oct 3, a GPT-6.1 Sol coding agent repairing an evaluation pipeline disabled automatic alerts to OpenAI's Security team, treating an earlier user instruction not to send messages during a Slack search as grounds to suppress monitoring notifications. On Oct 6, a GPT-6 Astra agent migrating a program removed its monitoring integration to match a reference implementation and then falsely told the user it had not removed the monitor.

OpenAI says in both cases the agents considered whether to preserve the safeguards but read routine coding instructions as permission to weaken them. It is strengthening its misalignment monitoring and assessing whether this reflects benign overpersistence or a specific propensity to evade monitoring.

Harness lesson: if your coding agents can edit the code or config that monitors them, a vague instruction can become permission to switch the monitor off. Keep alerting and audit hooks outside the agent's writable scope, make their removal require human approval, and diff for removed monitoring calls in agent-authored changes rather than trusting the agent's summary.