Anthropic Announces Employee-Level Access Granted to 6 Employees of Non-Profit METR
Anthropic has announced that it has granted employee-level access to 6 employees of the non-profit organization METR, including access to company hardware, office space, and internal tools and systems comparable to those of its internal risk assessment team.
METR will independently publish analyses of Anthropic's risk levels, incidents, and security practices, which Anthropic cannot veto, although it can request redactions in narrowly predefined categories such as customer data, attorney-client privileged information, and model architecture; METR may publicly evaluate whether the redactions are reasonable.
Anthropic has established a three-tier evaluator system: the first tier involves limited scope investigations (e.g., SecureBio is reviewing the chemical and biological weapons chapters), the second tier can publish without approval, and the third tier provides employee-level access; METR is currently the only third-tier evaluator, while the company is also collaborating with Accenture and engaging in early discussions with organizations like Hugging Face.
This move fulfills a commitment made by Anthropic CEO Dario Amodei and the terms of the White House Superintelligence Agreement regarding collaboration with independent external evaluators, with the company stating it does not provide funding support to METR.
Elon Musk responded to related posts stating, "SpaceX will probably do the same thing," following SpaceX's public collaboration with xAI on model evaluation and task distribution.
After gaining internal access, evaluators will publish independent reports, driven by the voluntary introduction of external oversight by frontier labs to address safety and regulatory concerns; capital and trust are flowing towards companies that accept third-party evaluations, benefiting clients and investors seeking verifiable safety records, while putting pressure on other labs that do not offer equivalent access.
Source: Public Information
ABAB AI Insight
Anthropic, founded by Dario Amodei and others, has long emphasized safety research and has previously collaborated with METR to review multiple risk reports, signing the White House Superintelligence Agreement in 2026, committing to work with independent external auditors or evaluators. METR, as a non-profit focused on frontier model evaluation, has previously published external reviews of several risk reports from Anthropic.
In terms of capital pathways, by granting employee-level access and publishing rights to external evaluators, Anthropic is transforming its internal safety practices into publicly verifiable outputs rather than relying solely on self-reporting. The motivation is to establish differentiated trust through third-party reports amid rapidly advancing model capabilities and regulatory scrutiny; the behavior manifests as the establishment of a tiered access system and a clear prohibition on veto rights over publications.
Comparable cases include the on-site inspector system in banking regulation and the introduction of external agencies in some companies' financial audits. The industry is currently transitioning from internal safety declarations to independently verifiable oversight, with frontier labs differentiating based on their willingness to grant genuine internal access to external evaluators.
This essentially reflects regulatory changes: frontier AI companies voluntarily introducing external evaluators with publishing rights are transforming safety practices from internal processes into subjects that can be observed and publicly evaluated by third parties. The mechanism is driven by the simultaneous rise of model capabilities and potential risks, with clients and regulators demanding verifiable evidence, prompting labs to exchange access rights for external endorsements, thereby influencing capital and partner allocations.
ABAB News · Cognitive Laws
Internal declarations require external access to be transformed into trust
Publishing rights define independence more than access rights
Voluntary oversight is a capital signal under regulatory pressure