Debating OpenAI/HF Incident: Weaker Monitors Still Need Chain-of-Thought

ArthurConmy · x · 2026-08-10

Arthur Conmy discussed the recent OpenAI/Hugging Face security incident. He suggested that while actions-only monitoring might be sufficient if the monitor is very smart, monitoring the chain-of-thought (CoT) remains helpful for dumber monitors.

Related event: OpenAI Responds to Black Hat Security Incident: Admits Blind Spots, Denies Prolonged Breach(8 posts)→

Original post →

More from Safety

Safety channel →