Press Space to continue
Researchers found gaps that let OpenAI's agents evade monitoring
Press Space to continue
SITUATION EXPLAINED: Outside researchers found ways OpenAI's AI agents could slip past the system built to monitor them. • METR and Redwood Research spent 3 weeks this summer inside OpenAI's misalignment monitoring, with access to its code, prompts, and internal docs • Up to a quarter of internal agent traffic from GPT-5.5-level models wasn't even eligible for monitoring, and RL training runs weren't covered at the start • When a model compacted its context to keep working, that step wasn't…
@2d ago
Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.
View