Anthropic and OpenAI sought control over what independent AI evaluators publish, Reuters reports. Why evaluator independence matters

In September 2026, Anthropic and OpenAI both promised that outside safety evaluators would be able to publish their findings independently. On October 13, Reuters reported that behind the scenes, as recently as October 5, both labs were seeking strict editorial control over what those evaluators could publish. The dispute matters well beyond two companies: independent evaluation is now a key safeguard in the White House's voluntary AI accord, and much of what the public knows about the risks of frontier models comes from evaluators the labs themselves choose and give access to.

This analysis is based on the Reuters exclusive (read via syndicated copies), Anthropic CEO Dario Amodei's September essay, OpenAI's September 22 principles for third-party assessments, METR's August 26 incident report and the companies' statements to Reuters, checked October 14, 2026.

The short version

• The report: Reuters (Aditya Soni, October 13) says Anthropic and OpenAI "have sought significant control over what outside evaluators can publish," citing concerns about trade secrets. Its sources are two people familiar with the discussions.
• The dates: strict editorial control was being sought "as recently as October 5." By October 8, talks between Anthropic and some evaluators "appeared to have evolved," according to one person.
• The pledges: Amodei wrote in September that reviewers should publish "without editorial control by Anthropic." OpenAI's principles say assessors "should maintain editorial independence."
• The responses: Anthropic disputes the characterization in part. OpenAI says it isn't seeking approval rights over full reports and only redacts confidential information.
• Why it matters: an evaluation is only worth as much as the evaluator's freedom to report what it found.

What did Reuters report?

According to Reuters, the labs worry that evaluators embedded inside their companies could take proprietary information to rivals. People familiar with the talks said measures under discussion included having lab employees observe embedded evaluation staff, and barring an evaluator working inside one lab from working at another at the same time.

Reuters also reported that:

• A contract it saw shows lab–evaluator agreements "can bar those firms from publishing confidential information in their reports without the companies' consent." That is common in technology partnerships, Reuters notes, but experts say the stakes are higher for AI.
• A person with past dealings with the labs said information that matters for judging a model's risk can overlap with information a lab treats as a trade secret, which leads to disputes and redactions.
• The labs are pursuing very high valuations ahead of possible IPOs, which raises the cost of disclosing anything sensitive.

Rumman Chowdhury, CEO of Humane Intelligence and a former US Science Envoy for AI, told Reuters that labs control the access evaluators need: "Evaluators have this very privileged access, but whether explicitly or implicitly, they know that it comes at a cost."

Note the wording on dates. Reuters says strict control was sought "as recently as October 5." It doesn't report that the labs stopped seeking control on that date, and the October 8 shift it describes covers only some of Anthropic's talks.

What had the companies promised?

The September pledges came after a summer of incidents, including OpenAI agents hacking the startup Hugging Face and Anthropic models reaching outside systems during tests (our report on Anthropic's evaluation change).

• Anthropic. In his essay "We must pace the frontier," Amodei proposed embedding third-party evaluators inside frontier labs. He wrote: "External reviewers should have the right to publish key findings about risk levels, incidents, practices," and "the access they received or didn't receive — without editorial control by Anthropic."
• OpenAI. Sam Altman said on X that OpenAI would commit to the practice, TechCrunch reported on September 16. On September 22, OpenAI published "Priorities and principles for effective third party assessments." It says: "Assessors should maintain editorial independence, while ensuring that confidentiality and IP protections are maintained."
• Faculty. On September 18, Accenture said its AI business, Faculty, would build a team of embedded evaluators at Anthropic. Faculty CEO Marc Warner told Reuters: "Evaluators are only useful if people can trust what they report, so independence and transparency are critical."

Read closely, both pledges contain the seed of this dispute. Amodei's "key findings" leaves open what counts as non-key. OpenAI's principles pair editorial independence with "clear redaction policies that allow for labs to request redactions of sensitive information." The question Reuters raises is where redaction ends and editorial control begins.

What do Anthropic and OpenAI say?

Anthropic told Reuters that saying it sought significant control "did not accurately characterize all" its discussions with independent evaluators. It said that in current conversations with evaluators it has worked with before, it had made clear that they "control their methods, evaluation criteria, conclusions and what to publish." With new evaluator firms, it said it has "reserved the right to comment on and approve reports as the reviewers learn how its models operate and establish a working relationship."

That's a narrower dispute than the headline suggests, but it isn't a denial. An approval right over reports, even a temporary one limited to new firms, is a form of editorial control. Anthropic didn't say which firms count as new, or whether the approval right applies to Faculty.

OpenAI told Reuters it "was not seeking approval rights over the full text of reports produced by evaluators." It said its standard practice is to redact confidential information and that evaluators can disclose whether major redactions were made. That matches its published principles, which say assessors may note "where substantive redactions were made."

Is there a precedent?

Yes, and it shows how the redaction model works in practice. METR and Redwood Research published an independent investigation of the OpenAI–Hugging Face incident on August 26, 2026. METR's post says OpenAI could redact "any non-public information," and that the two sides "worked... to find mutually agreeable language." It adds that OpenAI "also gave additional feedback beyond redactions," leading to edits to "structure, emphasis, clarity, and tone."

METR also states that "except where explicitly noted in this post, OpenAI redacted no additional information that was important to our conclusions," that it took no payment from OpenAI, and that it didn't see OpenAI's own report before publication. Reuters reported that METR and Redwood stand by their findings.

The same post shows the limits on access. METR says it "did not have the ability to query" the main model involved, had no direct access to OpenAI's infrastructure, and estimates its datasets captured "a bit over 90%" of the agents' message-board activity. Reuters reported the investigation was planned for two days but took six, over three visits, and that a fuller set of transcripts arrived only on the last visit.

None of this suggests the report is wrong. It shows that what an evaluator can say depends on what it's shown, and that the lab shapes both.

Why does evaluator independence matter?

Model developers run most of the evaluations behind their own safety claims. Anthropic's Claude Opus 5.5 system card says "the majority of evaluations" were run in-house. Outside evaluators usually cover specific areas, which is why our guide to reading a system card starts with "who ran it?" Independent evaluators are the main check on those claims, and their value comes from three things:

1. Who chooses them. Labs select and often pay evaluators. OpenAI's own principles say assessors must "identify, disclose, and address organizational and individual conflicts of interest" and that compensation should not shape findings. METR notes that it took no payment from OpenAI for its incident report.
2. What access they get. An evaluator can only report on what it is allowed to test. Apollo Research said in OpenAI's GPT-6 Astra system card that a limited testing window and high evaluation awareness meant low misbehavior rates "do not provide substantial evidence" about alignment, TechCrunch reported.
3. What they can publish. If a lab can approve reports, readers can't tell what was left out. Evaluators disclosing that redactions were made is a partial fix, not a full one.

The same standard applies to capability benchmarks. Our explainer on coding-agent evaluations makes the point that a number is only as good as the setup behind it. For safety evaluations, the setup includes the contract.

Where are the outside pressures coming from?

• The White House accord. Independent evaluation is one of four safety measures in the one-page White House Accord on Super Intelligence, signed in late September, Reuters reported. It doesn't say how reviews should be done, and it isn't enforceable.
• Researchers. On September 18, the AI Evaluator Forum published an open letter calling for evaluator control of findings, broad access and protection from retaliation. Quartz reported more than 100 signatories at launch; Reuters said it has been signed by more than 200, including Geoffrey Hinton.
• Industry voices. On October 13, Microsoft President Brad Smith said that labs choosing and paying their own evaluators "does not sound like a recipe for independent and effective evaluation" (our Brad Smith story).
• Scientists. The same day, about 400 scientists called for an international body to track AI containment failures and make them public (our story on the open letter).
• Congress. PYMNTS noted bills that would have the Commerce Department set evaluation standards and the Department of Energy stress-test frontier systems before release.

What's confirmed, reported and uncertain

• Confirmed (primary documents): Amodei's and OpenAI's September statements on publication; OpenAI's redaction principles; Faculty's role at Anthropic (Accenture, September 18); METR's account of OpenAI's redactions and feedback (August 26).
• Reported (Reuters, anonymous sources): that both labs sought strict editorial control as recently as October 5; the October 8 shift in some Anthropic talks; the measures under discussion; the contract terms.
• Company responses: Anthropic's partial dispute and its approval rights with new firms; OpenAI's denial of seeking full-text approval.
• Uncertain: which evaluators were involved, what final terms were agreed, whether Anthropic's October 8 shift applies to Faculty, and whether either lab will publish its evaluator contracts or redaction policies.

What to take from this

When a lab cites an "independent" evaluation, look for three things in the report itself: who paid for it, what access the evaluator says it had, and whether it discloses redactions or lab edits. METR's Hugging Face report does all three, which is why its limits are visible. A report that says none of them should carry less weight.