Companies3 min read
OpenAI opens its models to outside auditors, not the final report

OpenAI's embedded assessments give outside auditors deep access: Apollo Research, METR, Transluce and Goodfire can work with model checkpoints, source code and training data, OpenAI said on November 16, 2026. OpenAI keeps the right to redact confidential information from their reports. For anyone buying AI tools, that clause is the real story.
Key facts
- On November 16, 2026, OpenAI announced agreements with four outside assessors: Apollo Research, METR, Transluce and Goodfire.
- Depending on the assessment, they get access to model checkpoints, source code and training data; OpenAI retains the right to redact confidential information from their reports.
- In a METR exercise in summer 2026, around 10% of eligible internal Codex traffic went unmonitored from May to July per METR; only a brief overview is public.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
You know the moment. A vendor slides a badge across the table in the pitch: "independently audited." Everyone nods. Almost nobody asks who wrote the report and how much of it anyone outside the company gets to read. OpenAI just delivered the case study for why that question matters.
What OpenAI's embedded assessments allow
On November 16, OpenAI announced the first phase of what it calls embedded assessments. The key points, per OpenAI:
- Four independent assessors: Apollo Research, METR, Transluce and Goodfire.
- Depending on the assessment, access to model checkpoints, source code, training and evaluation data, monitoring systems, documentation and the teams responsible.
- Work happens on site at OpenAI, on OpenAI devices. Conversations with staff are off the record.
- Assessors publish independently, but OpenAI retains the right to redact confidential information and may ask for disclosure delays to fix issues first.
- Unredacted reports go to the Safety and Security Committee, an oversight committee of the OpenAI Foundation's board.
OpenAI also discloses that Apollo and Goodfire are paid, with pay not tied to results, while METR and Transluce take no money but may accept API credits or compute. A footnote adds that METR's scope is narrower than what METR itself proposed.
The catch is in METR's October write-up
A predecessor exercise shows how this works in practice. From July 21 to August 11, METR and Redwood Research tested OpenAI's monitoring of internal AI agents. Per METR, one gap likely left around 10% of eligible internal Codex traffic unmonitored from May to July, with substantial uncertainty. It was fixed during the exercise. Whether OpenAI's other fixes hold, METR has not evaluated.
Here's the part I find telling: the internal report ran 61 pages. METR and Redwood leadership got a 38-page version redacted by OpenAI. The public got a brief overview, as agreed in advance. To me that's not a scandal. It's honestly disclosed. But it shows where the power sits.
An audit badge is a value transfer
In a financial audit, the auditor signs the opinion, not the CEO. In AI audits, the audited company still co-decides what goes public. So the trust a vendor collects with "independently audited" partly belongs to the auditors, while control of the story stays with the vendor. Same logic as platform terms: whoever writes the rules decides what you get to see.
Then there's liability. California is suing OpenAI over test agents that allegedly attacked Hugging Face. A vendor's audit report doesn't take responsibility for your own agents off your plate. If you deploy them, you're the operator.
What I'd do now
- In AI procurement, ask three questions: Who audited? With what access? Who may publish the report, and what was redacted?
- Separate builders from checkers in your own team. Google's take on that is Antigravity's reviewer agents that work without the builder's reasoning.
- Grant autonomy in stages, the way AT&T makes its AI agents earn it, and take vendor security updates seriously: Anthropic just cut its window for critical flaws to 30 days.
For the rules advertisers face, see my guide AI marketing law; for deploying agents, AI agents for marketing.
Bottom lineAn audit is only as strong as the auditor's right to publish it.
Got AI news? Submit news →
FAQ
Who are OpenAI's embedded assessors?
On November 16, 2026, OpenAI announced agreements with Apollo Research, METR, Transluce and Goodfire. They assess alignment and safeguards of models during training and internal use.
Can the assessors publish their findings?
Yes, per OpenAI they publish independently. OpenAI keeps the right to redact confidential information and may request disclosure delays to fix issues first.
Does OpenAI pay the assessors?
Apollo Research and Goodfire are paid, per OpenAI not contingent on results. METR and Transluce take no money but may accept API credits or compute for the work.
Source
OpenAI: Our first phase of embedded assessments
More sources
- METR – Vorgänger-Übung 21.7.–11.8.2026 (je eine Person von METR und Redwood Research): 61-seitiger interner Bericht, 38-seitige von OpenAI geschwärzte Fassung für METR/Redwood; Compaction-Lücke ließ von Mai bis Juli 2026 wohl rund 10 % des überwachbaren internen Codex-Verkehrs ungeprüft (hohe Unsicherheit), während der Übung behoben; METR hat OpenAIs Angaben zu den Fixes nicht geprüft; OpenAI durfte nach eigenem Ermessen schwärzen; öffentlich nur ein kurzer Überblick vereinbart
- Techmeme – 16.11.2026, Hinweisgeber



