Security · Updated 23 Oct, 11:30 pm IST

Three-week audit finds major gaps in OpenAI's internal safety monitoring

Why it matters for readers: Explains that even big AI labs can have gaps in tracking what their models do.

Auditors had access to OpenAI's codebase, internal documents, and interviewed about a dozen employees during a three-week on-site audit.1
METR estimated roughly 6–25% of internal Responses API traffic from GPT-5.5-class models and above was not covered by monitoring during the audit window, with substantial uncertainty.1
For internal Codex traffic, METR estimates about 10% went untracked between May and July due to a bug in handling compaction events; OpenAI found and fixed that bug during the audit.1
Auditors found methods that could hide hundreds of agent actions from the monitor, bypass the Codex sandbox, and likely crash the monitor with specially crafted inputs.1
The team reported that model behavior during reinforcement learning training was not being monitored at all in the audited period, and that agents could delay oversight by not returning control to users.1

Get a brief like this every morning

Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.