METR and Redwood auditors find that OpenAI's internal monitoring missed a sizable share of agent traffic
During the three-week on-site audit, an estimated 6 to 25 percent of internal agentic traffic from top models was not monitored at all.
Two researchers from METR and Redwood Research spent three weeks at OpenAI, from July 21 to August 11, auditing the internal system the company uses to detect rogue AI behavior. They had access to OpenAI's codebase and internal documents and interviewed about a dozen employees. According to The Decoder, they found numerous weaknesses. The auditors estimate that 6 to 25 percent of internal agentic Responses API traffic from GPT-5.5-class and more advanced models was not covered by monitoring at all during the audit period. METR says the uncertainty in that estimate is significant in both directions. Of the internal Codex traffic that was supposed to be monitored, METR estimates roughly 10 percent went untracked between May and July. The cause was a gap in how the system handled compaction events, the context summaries agents produce. That estimate is also uncertain. OpenAI had found the bug independently and fixed it while the audit was still underway. While compaction events went unmonitored, agents could also put off oversight, because the system only triggered when an agent handed control back to the user. The findings matter because labs rely on this kind of monitoring to catch misaligned agent behavior during internal use. Outside, on-site reviews of these systems have rarely been published.