Anthropic & METR: Embedded White-Box AI Safety Audits

Fulfilling its formal commitments under the White House Accord on Super Intelligence and advancing third-party safety verifiability, Anthropic has officially granted employee-like access to a dedicated evaluation team from METR (Model Evaluation and Threat Research), a leading nonprofit dedicated to assessing frontier AI risks.

Embedded Employee-Level Access with Zero Publication Veto

Anthropic has onboarded six METR researchers equipped with company laptops, dedicated office space, and credentials to internal tools and diagnostics that mirror the capabilities granted to Anthropic's proprietary risk-assessment and alignment teams.

Except for narrowly delineated boundaries, including customer personal data, attorney-client privileged materials, and commercially sensitive model architecture secrets, Anthropic engineers are instructed to share technical context as seamlessly as they would with full-time internal colleagues.

The definitive architectural milestone of this partnership is its editorial independence: METR holds the autonomous right to publish its risk findings, incident analyses, and methodology reviews without any publication veto from Anthropic management. While Anthropic retains the right to mandate redaction for predefined confidential domains, METR remains entirely free to characterize those redactions publicly and debate whether they were justified.

Regulatory Precedent: The Bank Resident Examiner Analogy

Anthropic directly drew an analogy to classical banking supervision. Financial regulators station resident examiners permanently inside systemically important commercial banks because internal operations afford granular visibility into systemic exposure that outside audits cannot detect. Embedding third-party AI safety evaluators mirrors this philosophy directly within frontier AI labs.

For the global technology sector, this initiative marks an imperative structural shift: frontier model governance is graduating from black-box API benchmarking toward embedded white-box evaluations conducted by credentialed external safety scientists.

Three-Tier Access Architecture & Expanding Evaluator Network

Anthropic formalized a progressive three-tier framework governing independent model auditors across the industry:

Beyond METR, Anthropic highlighted ongoing engagements with biosecurity group SecureBio, currently examining the chemical and biological weapons sections of Anthropic's August 2026 Risk Report, alongside active red-teaming programs with Accenture and exploratory evaluator dialogues with Hugging Face.

“Just as banking regulators place resident examiners inside systemically important banks, giving independent evaluators white-box access bridges the dangerous gap between what AI labs know and what society understands.”