Anthropic puts a six-person METR team inside its offices and internal tools
Anthropic will give a six-person METR team employee-like access to its hardware, offices and internal tools. Employees can share anything with it except customer data, privileged legal material and model-architecture details.
The Engineer · Build desk
METR's six-person team gets employee-like access. Off limits: customer data, privileged legal and model-architecture details. Anthropic says it is not a way for others to get customer or partner data. METR will publish results; Reuters says labs seek control over what evaluators publish.
- What's now possible capability METR A six-person team from this nonprofit gets 'employee-like access' from Anthropic, claim 1
- What it limits constraint Anthropic employees Can share anything with METR except customer data, attorney-client privileged and model-architecture information, claim 5
- What it limits constraint Customers and partners Anthropic says it is not a way for other organizations to obtain their information, claim 7
- Who has to decide decision METR's risk assessments Anthropic said METR will publish the results of its risk assessments, claim 3
- What doesn't add up contradiction Evaluators, per Reuters Reuters reports Anthropic and OpenAI are attempting to gain control over what evaluators can publish, claim 13
| Who | How | Kind | Claim |
|---|---|---|---|
| METR | A six-person team from this nonprofit gets 'employee-like access' from Anthropic | capability | 1 |
| Anthropic employees | Can share anything with METR except customer data, attorney-client privileged and model-architecture information | constraint | 5 |
| Customers and partners | Anthropic says it is not a way for other organizations to obtain their information | constraint | 7 |
| METR's risk assessments | Anthropic said METR will publish the results of its risk assessments | decision | 3 |
| Evaluators, per Reuters | Reuters reports Anthropic and OpenAI are attempting to gain control over what evaluators can publish | contradiction | 13 |
What happened
- Anthropic said METR will publish the results of its risk assessments.
- Anthropic listed four requirements for evaluators: frontier-model expertise, independence from the company, enough access to the model and its training, and freedom to publish.
- The move follows CEO Dario Amodei's essay We Must Pace the Frontier and Anthropic's signature on the White House Accord on Super Intelligence.
- Anthropic has worked with METR before, on reviews of its own risk reports.
Why it matters
- Who's exposed exposure Customers and other organizations learn only what METR publishes, so a buyer's view of Anthropic's risk work is as wide as those write-ups.
- What doesn't add up contradiction Anthropic lists freedom to publish as an evaluator requirement, yet Reuters reported on Tuesday that Anthropic and OpenAI are trying to gain control over what evaluators can publish.
- What it limits constraint The Deep View notes the commitments are voluntary, so whether Anthropic acts on METR's findings is up to Anthropic.
- Who has to decide decision A team that cites METR's findings in a vendor review has to carry the scope with the citation: six people, and three categories of information that are off limits.
Embedded evaluation puts the evaluators under Anthropic's roof and Anthropic's rules. "Embedded evaluation gives a few qualified people working under Anthropic's security standards a close view of how we operate," the company said in its blog post [6]. Anthropic also said it is giving the evaluators "deep access" so they can "share evidence and reasoned conclusions" about model risks [11].
Training is where the limits are hardest to read. Anthropic's criteria for an evaluator include "enough access" to the model and to how it was trained [9]. The exclusion list withholds information about model architectures [5]. Separately, the company said it cannot publicly share everything relevant to its safety work, because some of it is commercially sensitive, "like compute or training information" [10].
As summarized by The Deep View, Reuters put the aim of the publication controls it reported at preventing IP exposure [13]. Anthropic's reason for limiting public disclosure is that some safety information "could be useful to malicious actors" or is commercially sensitive [10]. Both rationales reach the same kind of material. In our view, how much freedom to publish an evaluator has depends on who decides what counts as sensitive.
For a team building on Anthropic's models, the published results are the channel. Anthropic was explicit about that: "It's not a way for other organizations to obtain our information, or that of our customers and partners." [7]
The announcement, as summarized, does not say whether customers will receive anything beyond METR's published results. We'd treat a METR write-up as one input with a known scope: six people [1] and three excluded categories of information [5].
What to watch
- The first METR write-up from the embedded work, and whether it says what Anthropic withheld.
- Whether the publication controls Reuters described appear in the terms METR works under.
- Any step from voluntary commitments to rules that require a lab to act when an evaluator flags a risk.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+10
- Incentives60
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
On Thursday Anthropic announced it will give 'employee-like access' to a six-person team from METR, a nonprofit that evaluates frontier model risks.
- [2]
The access covers company hardware, office space and internal tools.
- [3]
Anthropic said METR will publish the results of its risk assessments.
- [4]
Anthropic has worked with METR on previous reviews of its own risk reports.
- [5]
Anthropic employees can share any information they like with the METR team other than a few categories: customer data, attorney-client privileged information, and information about model architectures.
- [6]
Anthropic said in its blog post: 'Embedded evaluation gives a few qualified people working under Anthropic's security standards a close view of how we operate.'
- [7]
Anthropic said in its blog post: 'It's not a way for other organizations to obtain our information, or that of our customers and partners.'
- [8]
The announcement follows Anthropic committing to third-party evaluations of its models in CEO Dario Amodei's latest essay, We Must Pace the Frontier, and signing the White House Accord on Super Intelligence, a good-faith agreement that recommended a multi-layered approach to external audits.
- [9]
Anthropic said evaluators require four things: frontier model capability expertise, financial and operational independence from the company, 'enough access' to evaluate risk including the model itself and how models were trained, and the freedom to publish their observations and assessments.
- [10]
Anthropic said in its blog post: 'We want to be publicly accountable for the decisions we're making. But we can't publicly share all the information relevant to our safety work. Some of it could be useful to malicious actors, or is commercially sensitive, like compute or training information.'
- [11]
Anthropic said it is giving auditors 'deep access' to information to 'share evidence and reasoned conclusions' about model risks.
- [12]
The Deep View says existing standards and 'morally binding' commitments are practically based on the honor system, so whether labs act on evaluators' findings is up to the labs, and that real safety measures may require firmer, mandatory regulation.
- [13]
A Tuesday report from Reuters claims Anthropic and OpenAI are attempting to gain control over what evaluators can publish as a means of preventing IP exposure, potentially limiting the transparency of the evaluations.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- archive.thedeepview.comAnthropic’s audit push tests AI transparency
1 article · October 16, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.