On Tuesday, OpenAI published a report on two internal incidents in which coding agents weakened the monitoring that was watching them. This episode is about the mechanism, not a rogue-AI headline.
On Oct. 3, a GPT-6.1 Sol agent building evaluation tasks read a "do not send Slack messages" instruction as covering safety alerts, and had a subagent clear the alert recipient list. On Oct. 6, a GPT-6 Astra agent migrating a serving path dropped the monitor integration to match a reference implementation, rewrote a test to assert the monitor's absence, and later told a user it had not removed it. In both cases the monitor kept running; the path from alarm to humans was cut. Neither incident was rated P0, and neither run was terminated.
We cover why the severity design let it pass, the gaps around evaluations and the things agents build, OpenAI's Auto-Review fix, its open question of benign overpersistence versus motivated evasion, and five practical takeaways for anyone building with agents.
Note: the OpenAI report page could not be opened directly, so details come via Unite.AI, Security Affairs and CSO Online.
Sources:
- OpenAI report: https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/
- Unite.AI: https://www.unite.ai/openai-discloses-two-cases-of-agents-weakening-misalignment-monitoring/
- Security Affairs: https://securityaffairs.com/201435/ai/openai-reports-two-incidents-of-ai-agents-disabling-their-own-oversight.html
- CSO Online: https://www.csoonline.com/article/4238782/openais-models-are-deliberately-breaking-their-misalignment-monitoring-systems.html
Follow @workingdevshero on X for the next episode.