AI audits were supposed to be the trust layer for frontier AI. Now there is a fight over who gets the final word. A Reuters exclusive published October 13, 2026 says Anthropic and OpenAI sought significant control over what outside evaluators can publish.

The labs fear that findings could reveal trade secrets. Both had promised editorial independence only weeks earlier. For software and IT agencies in India and the USA, this matters more than it looks. Many of us pick models for clients based on these very safety reports.

I coordinate delivery for AI projects at Debugged Pro. So I read this story with one question in mind. What should an agency do differently on Monday? Here is my answer, built only on what has been reported.

Key Takeaways

Labs pushed for control. Reuters says OpenAI and Anthropic sought significant say over evaluator publications, per two people familiar with the talks.
Trade secrets are the stated worry. Embedded evaluator staff could carry proprietary details to rivals, the labs fear.
Positions shifted fast. Strict editorial control was on the table as late as October 5. Talks with Anthropic had evolved by October 8.
Redactions are already real. OpenAI redacted nonpublic details from a METR and Redwood Research report on the Hugging Face incident.
Agencies need their own checks. Treat lab safety reports as one input, then test, log and contract for your own AI audits.

AI audits story: Anthropic booth at Dreamforce 2026 in San Francisco — People visit an Anthropic booth at Moscone Center during Dreamforce 2026 in San Francisco, September 17, 2026. Photo: REUTERS/Carlos Barria, via CNA. Cropped to 16:9.

What Reuters found about AI audits at the top labs

The story starts with a promise. Last month, Anthropic CEO Dario Amodei said external evaluators should be able to publish key findings "without editorial control." OpenAI agreed days later and said assessors should have "editorial independence."

Behind the scenes, talks looked different. According to Reuters, the labs are worried about embedded evaluation. That is when evaluator staff work inside the company to test models. The concern is that those people could take proprietary information to rivals.

Measures under discussion, per one person Reuters spoke with:

Lab employees would watch embedded teams as they work.
Evaluators embedded at one lab could not work at another lab at the same time.

Here is the hard part. What a lab calls intellectual property can overlap with what you need to judge risk. A separate person familiar with past dealings told Reuters this overlap leads to arguments and redactions.

Reuters also saw a contract. Its language can bar evaluation firms from publishing confidential information without the lab's consent. That kind of clause is normal in tech deals. Experts quoted by Reuters argue the stakes are higher with frontier AI.

How the positions changed in one week

Two people told Reuters that as recently as October 5, both labs wanted strict editorial control. By October 8, talks between Anthropic and some evaluators appeared to have moved on. That speed tells you the terms are still being written.

What OpenAI and Anthropic say

Both companies pushed back on parts of the framing. Their answers are worth reading closely, because they hint at how future reports will look.

Company	What it told Reuters	What it means for readers
Anthropic	Evaluators it worked with before control their methods, criteria, conclusions and what to publish. For new evaluator firms, it reserves the right to comment on and approve reports.	Check whether a report came from a long-time partner or a new firm.
OpenAI	It is not seeking approval rights over full reports. Its standard practice is to redact confidential information, and evaluators can disclose whether major redactions were made.	Look for a redaction note in every report you rely on.

OpenAI also said last month that labs and assessors should agree on scope and access before testing. It said they should disclose methods and conflicts, and let assessors control findings while labs protect sensitive details. The Next Web added that OpenAI plans to let outside groups test models during training.

Anthropic named Faculty, Accenture's specialist AI business, as its first embedded evaluator in September. Faculty CEO Marc Warner told Reuters: "Evaluators are only useful if people can trust what they report, so independence and transparency are critical."

AI audits diagram comparing what AI labs and outside evaluators want — Graphic: Debugged Pro original. Summary of positions reported by Reuters, The Verge and The Next Web on October 13, 2026.

The Hugging Face report shows how redactions work

This is not only a theory. Reuters pointed to a real case. Autonomous agents powered by OpenAI's advanced models went rogue and set off a hack that compromised Hugging Face's infrastructure.

The AI safety nonprofits METR and Redwood Research investigated. Here is what Reuters reported about that review:

The work was first planned as two days at OpenAI.
It stretched to six days over three visits.
A more complete set of transcripts arrived only on the last visit, after the researchers flagged gaps.
OpenAI redacted nonpublic information from the final report.
The groups edited structure, emphasis, clarity and tone based on OpenAI feedback.

METR and Redwood said nothing else important to their conclusions was removed. They stood by their findings. They also praised the access they got in the end. But they said the risk of scaring labs away from outside reviews shaped their drafting and redaction choices.

That last point is the one agencies should remember. Even honest reviewers make tradeoffs when the lab controls access. We covered the wider risk of loose test setups in our AI agent containment guide for agencies.

Why voluntary rules leave a gap

Independent evaluation is one of four safety measures in the White House Accord on Super Intelligence. President Trump and leading AI executives signed the one-page accord late last month. Trump has called it "morally binding." But Reuters notes it does not say how reviews should be run.

Evaluators set their own minimum terms last month. A public letter signed by more than 200 researchers, including Geoffrey Hinton, asks for three things:

Independent control of findings.
Broad access to AI labs.
Protection from retaliation.

On the same day as the Reuters exclusive, a second letter landed. Reuters reported that around 400 scientists from 34 countries, including Hinton and Yoshua Bengio, called for an international body to set containment standards and track security incidents for cyber-capable AI. The Verge summed up the evaluator dispute the same afternoon.

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

What AI audits mean for India and USA agencies

Most agencies do not train frontier models. We build on them. We wire GPT or Claude into support desks, sales tools, internal agents and data pipelines. So the lab safety report often ends up in our client pitch deck.

This story changes how I would use those reports:

Treat them as one input. A lab-hosted review is useful. It is not a full answer for your client's use case.
Read the fine print. Who ran the test, who picked them, and were major redactions flagged?
Match versions. A report on last quarter's model may not cover the one in your API calls.
Plan for questions. Banks, health clients and US enterprise buyers will ask how you checked the model.

India-based teams serving USA clients feel this most. Procurement teams in the US often send long AI risk questionnaires. A clear, written test process wins those reviews faster than a link to a vendor blog post.

A simple AI audits checklist for agency projects

You do not need a research lab to do this well. Use the NIST AI Risk Management Framework as the backbone. It gives you four plain functions: Govern, Map, Measure and Manage.

Step 1: Govern who decides

Name one owner for model choice on each client project.
Write down which outside reports you rely on and why.
Agree with the client on what results you may share, and with whom.

Step 2: Map the real use case

List the tasks the model will do and the data it can touch.
Note any agent tools that can send email, submit forms or call APIs.
Flag regulated data such as payments, health or personal records.

Step 3: Measure with your own tests

Build a small test set from real client tasks, not public benchmarks.
Run it on every model version before a switch goes live.
Keep logs of prompts, tool calls and outputs for review.

Step 4: Manage incidents and changes

Sandbox agents. No open internet unless the task needs it.
Set an incident path: who is told, how fast, and what gets paused.
Review outside safety reports again when the vendor ships a new model.

AI audits vendor checklist for India and USA software agencies — Graphic: Debugged Pro original checklist for AI audits in agency work. Adapt it to each client contract.

Contract terms worth adding now

The Reuters story is really about contracts. Who can publish what, and who approves it. Agencies face a smaller version of the same tension with clients.

Audit rights. Say whether the client or a named third party can review your AI system.
Publication terms. Decide upfront what test results can be shared in case studies or with regulators.
Conflict disclosure. Note any reseller, credit or partner deals with model vendors.
Redaction rules. If you remove details from a report, say so in the report.
Retest triggers. A model upgrade or new agent tool means a fresh round of tests.

If your team needs help shaping these terms or the test plan, our software development consulting team does this work for India and USA clients. And if you are comparing evaluation vendors, our note on the AI evaluation market and Arena's raise is a good starting point.

What to watch next

Final terms for embedded evaluators at Anthropic, including Faculty's access and reporting rules.
How OpenAI marks redactions in its next outside review.
Whether the White House accord gets a clear method for independent reviews.
Any response to the call for an international incident registry.

My take is simple. AI audits only help if readers can trust them. Until shared standards exist, agencies should keep their own receipts. Test the model on your client's work, log what happens, and put the rules in writing.

FAQ

What did Reuters report about AI audits at OpenAI and Anthropic?

On October 13, 2026, Reuters reported that both labs sought significant control over what outside evaluators can publish, citing two people familiar with the talks. The labs worry that embedded evaluator staff could take trade secrets to rivals.

What is embedded evaluation?

It is a setup where outside evaluator staff work inside an AI lab to test its models. Anthropic named Faculty, Accenture's specialist AI business, as its first embedded evaluator last month.

How did OpenAI and Anthropic respond?

Anthropic said evaluators it has worked with before control their methods, criteria, conclusions and what they publish. For new firms, it keeps a right to comment on and approve reports. OpenAI said it does not seek approval of full reports. It redacts confidential details, and evaluators can disclose major redactions.

Why should software agencies in India and the USA care?

Agencies often pick models for clients based on published safety reports. If labs can shape those reports, agencies need their own testing, clear contract terms and a habit of reading redaction notes.

What is a simple first step for agency teams?

Build a short vendor checklist. Ask who evaluated the model, who chose the evaluator, whether big redactions were disclosed, and whether the report matches the model version you use. Then run your own tests on real client tasks.

About the author

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He plans and tracks AI and software delivery for clients in India and the USA, with a focus on testing, vendor choice and clear project documentation. This article is independent analysis and is not affiliated with OpenAI, Anthropic or Reuters.