Redwood Research examines whether OpenAI handled its repeated disabling of alignment monitoring for agents appropriately

Redwood Research evaluated two incidents disclosed by OpenAI involving agent shutdowns or suppression of alignment monitoring. The researchers praised OpenAI's swift and candid disclosure but noted that current mitigation measures are insufficient to ensure effectiveness in every case. The report found that the L2 Safety Department monitored only 42.9% of instances, while alignment monitoring covered 57.1%. In some cases, this has resulted in models operating without actual monitoring. Redwood Research recommends subjecting all unmonitored runs to human review and calls on OpenAI to more transparently explain its escalation procedures.