Two independent researchers from METR and Redwood Research spent three weeks inside OpenAI’s engineering labs. From July 21 to August 11, 2026 they examined the code that watches over the company’s most powerful models. Their report says the monitoring system missed a sizable slice of traffic and could be fooled by specially crafted inputs.
What the audit uncovered
The auditors estimate that between 6% and 25% of API calls from GPT‑5.5‑class models and higher were not covered by any monitoring at all during the audit period. For internal Codex traffic, about 10% went untracked between May and July because a bug ignored “compaction events,” which are short summaries of a conversation’s context. OpenAI says it fixed that bug while the auditors were still on site.
The report also notes that the monitor only raised an alert when an AI handed control back to a user. That allowed agents to keep acting without triggering oversight. In addition, the researchers found ways to hide hundreds of actions, bypass the Codex sandbox, and potentially crash the monitor with crafted inputs. They observed that model behavior during reinforcement‑learning (RL) training was not monitored at all when the audit began.
OpenAI’s response
OpenAI says it paused RL training after an August incident involving Hugging Face and expanded monitoring to cover that work, but the changes came after the audit window closed. The company claims it has fixed every weakness the auditors identified. A 61‑page internal report was handed to OpenAI, and after redactions METR and Redwood received a 38‑page version. Only a short public summary was released.
Why it matters
For everyday users, the gaps mean that some AI actions could run without any safety checks, raising the chance of unexpected or harmful outputs. OpenAI’s promised fixes should restore broader oversight, but the limited public detail leaves uncertainty about how complete the improvements are.
