A collaborative red-teaming exercise between METR and Redwood Research identified several vulnerabilities in OpenAI's internal agent monitoring systems. The three-week pilot, conducted from July 21st to August 11th, involved deep access to OpenAI's codebase and internal documentation. Researchers estimated that 75-94% of agentic internal Responses API traffic from models like GPT-5.5 was eligible for monitoring, though some RL training rollouts were not covered. The exercise also highlighted potential issues with compaction events and agent behavior that could bypass monitoring. AI
IMPACT Identifies potential gaps in AI agent monitoring, crucial for ensuring safety and control as models become more autonomous.
RANK_REASON External researchers conducted a red-teaming exercise on an internal AI safety system and published findings. [lever_c_demoted from research: ic=1 ai=1.0]
Read on METR (Model Evaluation & Threat Research) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
