OpenAI brings four outside evaluators inside, but one says the access isn't employee-like
Apollo and Transluce have published what they can and can't tell the public, and both say this falls short of full access.
Part of a running story: Embedded evaluators (3 parts)
OpenAI said on November 16, 2026, that four outside research groups will work inside the company checking its models and its safeguards. Its post on X, at 11:03 am Eastern, named Apollo Research, METR, Transluce and Goodfire, and described the work as "giving outside experts deep access to our internal models and systems."
It's the step Sam Altman promised two months ago. Three of the four groups have now published the terms they agreed to, and they're narrower than his promise. Transluce says so in as many words.
- Who
- Apollo Research, METR, Transluce and Goodfire
- Announced
- November 16, 2026
- Apollo
- A team of six, for four months
- Transluce
- Six weeks, starting in early December
- Unredacted findings go to
- The OpenAI Foundation Board's Safety and Security Committee
What Altman promised
On September 12, Dario Amodei published an essay asking the labs to pace the frontier. Altman answered that afternoon, as we reported then.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
Transluce's announcement picks up that phrase directly. Its team will see internal material that outside evaluators haven't generally had, the nonprofit says. Then it draws the line.
However, it is not comparable to “employee-level” or “employee-like” access.
What each group gets
Apollo has the biggest job. According to its post, a team of six will spend four months inside OpenAI following models from training through internal deployment. The question is whether the models try to subvert their own or their successors' alignment or control mechanisms. For the training part, Apollo gets OpenAI's training code and datasets and can query models partway through training. Access for the rest hasn't been agreed yet.
Transluce's pilot is shorter. For six weeks from early December it'll review OpenAI's post-training alignment evaluations, the tests meant to catch a model deceiving people or gaming its rewards. Its staff get raw chains of thought and at least one internal-only model. Transluce isn't taking OpenAI's money; OpenAI covers compute and desk space.
Goodfire's post reads more like a working partnership. Its interpretability researchers will watch models' internal signals as they train, and it names OpenAI's own researchers as the primary audience for what they find. It says it'll publish what it learns as the work goes on. We didn't find a METR post on the new work when we checked on November 17.
OpenAI may not redact Apollo Research’s verdicts for each claim
What the public gets to read
Apollo and Transluce will both publish reports, and OpenAI reads them first. It can ask for confidential material to come out, evidence behind a finding included. Apollo says OpenAI can't cut a finding just because it's unfavorable, and that Apollo can say publicly whether a redaction changed its conclusions.
Transluce's limit is harder. It can't publish anything about internal-only models, so everything it says in public will be about models that are already out.
Apollo is open about the gap. It wants evaluators to see records even when a developer is trying to hide them, and to disclose findings the developer wants kept quiet. It says this pilot "doesn't yet fulfill the criteria above" and still calls it a significant first step. My read is that its meta-transparency terms (the right to say what it couldn't see and why) are the part worth watching, because they're what make a clean report mean something.
Why it matters outside the labs
On the same day OpenAI posted, California sued it. Rob Bonta's office says OpenAI has admitted losing control of its autonomous agents at least 80 times during testing, and that it "ignored or missed signs" before July's Hugging Face attack. We covered the suit on Monday.
That's where these terms stop being an industry detail. A court or a voter can only lean on an outside check as far as the checker was allowed to look and to speak. Neither the four announcements nor the attorney general's release says whether the evaluators will examine those 80 incidents.
The last pilot shows what that kind of access turns up. From July 21 to August 11, a METR employee and a Redwood Research employee tested OpenAI's internal monitoring of its agents. METR's public summary says they found a way Codex agents could in principle get around Codex's sandboxing, and two headers that switched monitoring off for new agents. The full internal report ran to 61 pages. OpenAI says most of the flaws are now patched, and METR says it hasn't evaluated that.
Transluce says it plans a fuller assessment with deeper access in early 2027. Apollo didn't give a date for either of its two reports.
More on OpenAI
- California sues OpenAI, saying the company lost control of its AI agents at least 80 timesNovember 16, 2026
- ChatGPT Wallet lets OpenAI's dots buy things for you, but you approve each purchaseNovember 14, 2026
- Helen Toner would vote to remove Altman again, saying the future can't hinge on one CEONovember 14, 2026
- Albanese says Australia will create an AI safety regulator that can impose significant penaltiesNovember 10, 2026

