AI
Outside auditors find holes in OpenAI's agent monitoring
Two researchers from METR and Redwood Research spent three weeks inside OpenAI and concluded that between 6% and 25% of internal agent traffic from GPT-5.5-class models and above went unmonitored during the audit. They also estimated about a tenth of supposedly watched Codex traffic slipped through a bug in how the system handled context compaction, which OpenAI found and fixed mid-audit. The 61-page report was redacted to 38 pages for the auditors; the public received a summary.