Two researchers from METR and Redwood Research conducted a three-week on-site audit of OpenAI’s internal monitoring system for agentic behavior between July and August. The audit uncovered notable tracking gaps, estimating that 6 to 25 percent of internal agentic traffic for models in the GPT-5.5 class went unmonitored during the period. Additionally, roughly 10 percent of internal Codex traffic was untracked due to a bug in compaction event processing, and model behavior during reinforcement learning training was unmonitored when the evaluation began.
OpenAI independently discovered and fixed the compaction bug during the audit and stated that it expanded monitoring to reinforcement learning training following an August incident on Hugging Face. Although OpenAI reports that all identified weaknesses have been resolved, full details remain redacted in the public summary released by the auditors.
Why it matters
Reveals that frontier AI labs face technical challenges maintaining continuous safety monitoring across complex agentic workflows.
Highlights growing pressure from independent safety groups for direct infrastructure access and public reporting on safety guardrails.
Source: the-decoder.com



