Oossa

OpenAI lets four auditors examine its AI systems on site

OpenAI gave Apollo Research, METR, Goodfire and Transluce direct access to unreleased models and data, marking the first on‑site safety review.

By Published by Oossa: 1 min read

Christina @ wocintechchat.com M · Unsplash

OpenAI announced that four independent groups will work inside its offices to audit its AI systems. The auditors – Apollo Research, METR, Goodfire and Transluce – will see source code, training data, monitoring tools and unreleased models. They can also interview developers in confidence. OpenAI says the pilot follows commitments made under the White House’s “Accord on Super Intelligence.”

Apollo’s focus is on whether models try to subvert their own alignment or that of future versions. METR will look for dangerous behavior that OpenAI’s own tests might miss and will also test OpenAI’s ability to catch such behavior. Goodfire will use its interpretability platform to map internal model processes and suggest fixes. Transluce will evaluate the post‑training safety checks and plans an eight‑week follow‑up review in early 2027. An independent safety board will receive the full, unredacted reports and mediate any disputes. OpenAI can redact confidential material and delay public release to address risks first.

The program builds on an earlier on‑site review of OpenAI’s monitoring systems. It also expands a practice that started at Anthropic, where METR researchers were given employee‑level access last month. Critics note that Apollo and Goodfire are being paid for their work, while METR and Transluce receive only compute or API credits. Some safety experts, such as Redwood’s CEO Buck Shlegeris, say the access granted does not yet meet the “employee‑like” level many have called for.

Why this matters

For anyone using ChatGPT or other OpenAI products, the audits could surface hidden failure modes before the models reach the public. If the auditors find serious safety gaps, OpenAI may have to patch or delay releases, which could affect feature rollouts and pricing.

Why it matters

The audits give an outside view of how OpenAI’s most powerful models behave during training, not just after release. That could lead to safer products for everyday users, but the limited public disclosure means outsiders cannot fully verify the findings yet.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.