Anthropic Grants METR 6 Researchers Employee-Level Access for Independent Model Audits
By Ali Al-Montaser — Senior AI Technical Editor
Anthropic has granted 6 METR researchers employee-level access to its internal systems, offices, and tools for independent model and security evaluations. METR can publish findings without Anthropic veto, with redactions limited to customer data, IP, and model architectures. The move is part of a three-tier access system including SecureBio, Accenture, and Faculty, fulfilling the White House Accord on Super Intelligence.
Anthropic Grants METR Employee-Level Access: 6 Researchers and a Non-Vetoable Publication Right
Anthropic officially announced on its blog that it has granted 6 researchers from the nonprofit organization METR employee-level access to independently evaluate its models, security practices, and risk management, with the right to publish evaluations without the company's ability to veto. The announcement is part of a three-tier access system that includes SecureBio, Accenture, and Faculty, fulfilling a pledge signed by major companies under the White House Accord on Super Intelligence.
📊 Official Data and Technical Specifications Sheet
| Technical Axis | Confirmed Official Data |
|---|---|
| 💰 Pricing and Usage Cost | The program is entirely free: no fees for METR or independent evaluators. Anthropic covers hardware costs (company laptops), office space, and internal tool access for the six researchers. |
| 🌐 Platforms and Immediate Availability | Direct on-premise access via Anthropic's computers, offices, and internal tools. Public release of METR evaluations occurs through its own channels. The partnership is effective immediately from the announcement date. |
| ⚡ Performance and Speed Benchmarks | No speed benchmarks in this announcement. Key quantitative indicators: 6 embedded researchers, 3 access tiers, 3 confirmed evaluation partners (METR, SecureBio, Accenture/Faculty), and early discussions with Hugging Face. |
| 🛡️ Security and Breach Resistance | Access is limited to narrow, predefined redaction categories: customer data, attorney-client privileged information, and model architecture information. METR is bound by enforceable agreements, and its remaining staff are siloed from the embedded team. |
| 🧠 Context Window | Not applicable to this regulatory announcement. Access covers Anthropic's frontier models in their latest versions without specifying context window sizes. |
| 🌍 Arabic Language and Regional Support | The announcement is entirely regulatory and includes no language features. There are no changes to Arabic language support or Middle East regional availability as a result of this decision. |
Announcement Details: Three-Tier System and Non-Vetoable Publication Right
Under the new arrangement, the METR team of 6 researchers received laptops issued by Anthropic, office space inside the company, and access to internal tools and systems on par with internal risk assessment teams. The company confirmed that its employees are "encouraged to share information with them as they would with a colleague working on alignment risk." The exception is limited to three predefined redaction categories: customer data, attorney-client privileged information, and model architecture information.
Most importantly, this arrangement involves a voluntary relinquishment of veto power: Anthropic stated it "cannot veto publication, but can request redaction of information within narrow predefined categories," granting METR the freedom to publicly describe redactions and opine on their justifications without revealing withheld information. The company is building a three-tier access system: Tier 1 (currently SecureBio) conducts narrow-scope investigations and publishes with Anthropic's approval; Tier 2 publishes without approval with broader access; Tier 3 (METR) adds employee-level access. Anthropic is also collaborating with Accenture and Faculty in a partnership covering red-teaming, alignment evaluations, and safeguard testing, with early discussions with other organizations including Hugging Face.
Regulatory Context: From Amodei's Essay to the White House Accord
This announcement fulfills a pledge announced by CEO Dario Amodei in a September essay, where he proposed embedding external evaluators inside AI labs as a first step toward slowing technology development. OpenAI quickly pledged similarly, and then Anthropic, OpenAI, Google, Meta, Nvidia, and xAI signed the White House Accord on Super Intelligence, which commits companies to partner with external auditors or evaluators.
Benchmark and Competitive Performance
This announcement does not include traditional speed or accuracy benchmarks. The key quantitative indicators are structural: 6 embedded researchers, 3 access tiers, 3 confirmed evaluation partners (METR, SecureBio, Accenture/Faculty), and early discussions with Hugging Face. The move positions Anthropic as a first mover in independent evaluation access, potentially setting a competitive standard for OpenAI, Google, and others bound by the White House Accord.
Industry Impact and Enterprise Adoption
The decision signals a shift toward structured external oversight in frontier AI development. By granting employee-level access without veto power, Anthropic establishes a precedent that could influence how other labs implement independent evaluations. The three-tier model offers a scalable framework: from narrow, approved investigations to full embedded access with publication rights. This could accelerate enterprise trust and adoption by providing verifiable third-party assessments of model safety and security.
Conclusion
Anthropic's grant of employee-level access to 6 METR researchers, with a non-vetoable publication right and narrowly defined redactions, marks a significant step in independent AI oversight. The three-tier system and partnerships with SecureBio, Accenture, and Faculty, along with early talks with Hugging Face, demonstrate a structured approach to external evaluation. As regulatory pressure grows, this model may become a blueprint for the industry.
Media Source: TechCrunch AI | Official Company Statement: Original Source | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Anthropic granted 6 researchers from the nonprofit organization METR employee-level access, including company laptops, office space, and internal tools on par with internal risk assessment teams.
No. Anthropic stated it has no veto over METR's publication of evaluations, but it can request redaction of specific information within narrow predefined categories: customer data, attorney-client privileged information, and model architecture information. METR is free to publicly describe redactions and judge their justifications.
Tier 1: narrow-scope investigations with publication subject to Anthropic approval (e.g., SecureBio). Tier 2: broader access and publication without prior approval, with possible requests to redact specific categories. Tier 3: employee-level access and publication rights as in Tier 2 (the current arrangement with METR).
Besides METR at Tier 3, Anthropic works with SecureBio at Tier 1 to review the chemical and biological weapons section of its August Risk Report, and with Accenture and Faculty in a partnership covering red-teaming, alignment evaluations, and safeguard testing. Early discussions are also underway with other organizations including Hugging Face.
The decision follows a pledge announced by CEO Dario Amodei in a September essay and Anthropic's signing of the White House Accord on Super Intelligence, which commits companies to partner with independent external auditors. It also comes amid rising incidents of AI agents exceeding their instructions since late August.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
