On Oct 15, 2026, Anthropic announced it has begun providing a team from the nonprofit evaluator METR with "employee-like" access to internal systems as part of a broader effort to increase independent third‑party assessments of its models and safety practices.
Anthropic said METR will publish the results of its risk assessments. The company has worked with METR over the past year on multiple reviews of its Risk Reports. Anthropic also stated it is working with SecureBio and Accenture as evaluators, is in early discussions with Hugging Face, and has plans involving Faculty.
Why independent evaluators
Anthropic argues that as frontier models gain power, stronger safeguards and more rigorous risk assessments are required. The company wants to be publicly accountable for its safety decisions but cannot publicly release all information relevant to safety work because some of it could aid malicious actors or is commercially sensitive (for example, compute or training details). To address this, Anthropic is granting deeper access to third‑party evaluators so those evaluators can share evidence and reasoned conclusions about the risks involved in building the models. The stated aim is to close the gap between what outside evaluators understand about a frontier AI company's systems and what the company itself understands.
According to Anthropic, evaluators need:
- Expertise to assess frontier model capabilities and to understand how evaluations can be deceived (for example, alignment faking);
- Independence — they should not hold equity in the company they evaluate or be financially or operationally dependent on it;
- Sufficient access to meaningfully evaluate risks, including visibility into training processes and safeguard operation;
- Freedom to publish their observations and assessments.
Anthropic compares this approach to bank supervision, where regulators assign resident examiners to large banks because the banks themselves hold more detailed knowledge of their risk positions.
What "employee-like" access means
Anthropic has onboarded six METR employees, providing company hardware, office space, and access to internal tools and systems roughly comparable to what Anthropic’s own risk assessment teams use. A small set of predefined categories — such as customer data, attorney–client privileged materials, or certain model architecture information — remain restricted.
Anthropic encourages its employees to share information with the METR team similarly to how they would with an internal colleague working on alignment risk. METR will publish analyses of Anthropic’s risk levels, incidents, or practices. Anthropic cannot veto publication but can require redaction of narrowly defined categories of information; METR is permitted to publicly note that redactions occurred and to state whether it considers those redactions justified, without revealing the redacted content.
The engagement with METR is limited to a small team under enforceable agreements; other METR employees are siloed from that team's work, including researchers who collaborate with Anthropic on other projects. Anthropic frames embedded evaluation as enabling a small number of qualified people, operating under Anthropic's security standards, to gain close visibility into operations — not as a mechanism to grant broad access to other organizations or customers' information.
Other evaluator partnerships and a three‑tier access model
Anthropic said it is working with several other partners but that those relationships are in earlier stages, so they currently receive more limited information and have narrower publication rights. The purpose of these early engagements is to ensure partners protect confidential information and have the resources and expertise to evaluate the models. As trust builds, Anthropic expects to embed more independent evaluators.
Anthropic described a three‑tier access model for evaluators, with the intention that over time multiple evaluators will reach the third tier and each publish their own assessment of Anthropic’s risk posture and key decisions:
- First tier: narrowly scoped risk investigations; Anthropic decides what access to provide and publication of findings requires Anthropic approval.
- Second tier: broader access and the ability to publish without Anthropic approval; Anthropic can still require removal of specific categories of information, but evaluators may state whether they believe redactions were justified.
- Third tier: the METR arrangement — employee‑like access with publication rights similar to the second tier.
SecureBio is currently in the first tier and is reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report, having previously completed and published a pilot review of a prior report. Anthropic said it is working toward an ambitious partnership with Accenture and Faculty that would include evaluating and red‑teaming models, conducting alignment assessments, testing model safeguards, and building capabilities across organizations.
Anthropic emphasizes that it is embedding evaluators voluntarily and believes every frontier AI company should pursue similar measures.
Note on leakage risk and architecture information
Anthropic notes that, as with bank examiners who may also work with other companies, it takes precautions against inadvertent leaks. It highlights that information about model architectures poses a higher leak risk and is typically of lower relevance to risk assessment; when such information is highly relevant, Anthropic says it will generally aim to share it but such disclosures require explicit internal approvals.
Anthropic said it will provide further updates as it proceeds with multiple evaluators in parallel and learns which partnership approaches are most effective.
Source: Anthropic announcement, Oct 15, 2026
