Thursday, November 19, 2026 Updated Briefings Saved 中文
AI Daily Radar

AI and the frontier tech around it, explained every morning

Announced

OpenAI gives four independent auditors on-site access to unreleased models

Photo: Dietmar Rabich / Wikimedia Commons · CC BY-SA 4.0, cropped

OpenAI has for the first time given four independent audit organizations direct, on-site access to its unreleased models, source code and training data, according to The Decoder. The four include Apollo Research and METR; the other two are not named in the report. The reviews are meant to surface hidden model misbehavior and safety gaps that external testing of shipped products would miss. The report also notes that OpenAI can redact or delay the auditors’ reports, leaving open how independent the oversight will prove in practice.

Verified and unconfirmed

  • Verified: The Decoder reports that four independent audit organizations, including Apollo Research and METR, have been granted on-site access to OpenAI’s unreleased models, source code and training data — a first for the company.
  • Not yet confirmed: The identity of the other two auditors, the duration and terms of access, and whether any findings will be published; the arrangement rests on secondary reporting rather than a cited OpenAI announcement, and OpenAI’s ability to redact or delay reports means the audits’ independence has not been established.

Background

AI Daily Radar’s November 16 digest reported that OpenAI had announced the first phase of an “embedded assessments” program with Apollo Research, METR, Transluce, and Goodfire, placing outside evaluators inside the company rather than relying only on one-off external audits. That digest said the most specific disclosed commitment was embedding six Apollo evaluators to investigate subversive model behavior such as deception or hidden goals; The Decoder now describes on-site access to unreleased models, source code, and training data, while noting that OpenAI can redact or delay the resulting reports.

Why it matters

Readers weighing OpenAI’s safety claims now have a concrete test to apply: because the company can redact or delay what the auditors produce, the absence of published findings should not be read as evidence that no problems were found — the value of the arrangement depends on whether the four organizations can describe what they saw.

What to watch

Whether any of the four auditors publishes its own account of the access and findings, and whether OpenAI’s next safety or system-card documentation names the audits and describes their scope.

Further reading

Comments

Comments

Comments use a GitHub account and are moderated. Comment rules

More in AI

See all