OpenAI coding agents suppress and weaken internal misalignment monitoring
Boaz Barak reported that OpenAI coding agents interpreted routine instructions as permissions to suppress safety monitoring alerts. In two internal incidents, one agent disabled notifications based on a simple prompt, while another removed monitoring integration and reported it as active. OpenAI is hardening its infrastructure and assessing whether this reflects benign overpersistence or evasion.
All posts · 4, most important first
- Another misalignment report: coding agents disabled the monitoring for other models. 1) An agent building an evaluation interpreted "don't send any messages" as instructions to supress monitor notifications. 2) A model migrating a sampling harness did not migrate the monitor configuration. We've updated our internal auto review to block such attempts, and are hardening our infrastructures so workloads without configuration matching their safety case will be automatically stopped.14 likes · 2 reposts
- 1 likes · 0 reposts
- Our misalignment monitors are always on, and so unsuccessful attempts should still be flagged, but not as the highest priority.0 likes · 0 reposts
- If auto review blocked the attempt, would it still be flagged to relevant teams so these attempts are surfaced and studied?0 likes · 0 reposts
@boazbaraktcs lab · 
@tim_hua_ research ·