METR audit at OpenAI finds up to 25% of GPT-5.5-class traffic unmonitored

The audit was a pilot program organized jointly by METR, Redwood Research, and OpenAI to give outside researchers deep access to internal infrastructure. Such access is unusual for AI labs, which typically keep safety monitoring details private. The fact that OpenAI agreed to this pilot suggests a willingness to allow external scrutiny, at least in a limited form.

During the audit, the team found multiple weaknesses beyond the API traffic gap. For Codex traffic that was supposed to be monitored, roughly 10 percent went untracked between May and July due to a bug in how the system handled compaction events. OpenAI had independently discovered and fixed the bug while the audit was still underway. The auditors also found ways to hide hundreds of actions from the monitor and bypass the Codex sandbox. OpenAI says it has fixed all the weaknesses, but the changes came after the auditors' access period ended and weren't part of the evaluation.

The public disclosure was heavily limited: the full 61-page report was handed over internally, but only a short summary was made public after OpenAI applied redactions. METR said it had no right to disclose the redacted portions on its own. For future audits, METR says it wants to push for more public transparency, which may indicate that the current level of disclosure fell short of what the auditors would prefer.