Anthropic Opens Lab to Outside Auditors in Push for AI Safety Oversight
The AI maker is granting a nonprofit research group deep access to its systems and models, positioning embedded evaluators as a new standard for industry accountability.
Anthropic has opened its doors to external auditors with what it describes as "employee-like access" to evaluate safety risks in its AI systems, acting on earlier commitments to strengthen protections against potential harms from increasingly powerful models.
METR, a nonprofit organization that has previously collaborated with OpenAI, Anthropic, Google and Meta, has now stationed a team inside Anthropic's offices. These evaluators have been granted access to the company's hardware infrastructure and internal systems.
Third-party evaluation represents one of the most significant concrete measures that major AI labs have committed to implementing in response to mounting concerns about the potential consequences of advanced AI systems on economic stability, cybersecurity and broader societal risks.
In September, Anthropic chief executive Dario Amodei secured an unusual alignment between industry competitors when OpenAI's Sam Altman and SpaceX's Elon Musk both endorsed his proposal to station independent auditors within leading AI laboratories. Thursday's announcement represents the initial realization of these commitments. OpenAI has not yet announced its own plans for bringing in evaluators, despite Altman's earlier support for Amodei's concept.
As models have become more powerful, they've required stronger safeguards and more challenging risk assessment. 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.
Anthropic
Some observers have raised concerns that embedded evaluators could create conflicts of interest, drawing parallels to problems that have emerged in financial auditing, and have called for stronger regulatory intervention from government authorities.
Anthropic countered that releasing sensitive information publicly could aid bad actors or provide advantages to competitors. The company framed its evaluator model as analogous to banking regulation, where supervisory agencies place resident examiners within major financial institutions. "If a bank's risk-taking could threaten the public, third parties try to see that in advance," Anthropic stated.
The company is also navigating investor expectations as it prepares for an initial public offering anticipated later this year at a valuation exceeding $2 trillion. Safety considerations have become a focal point in Anthropic's investor discussions this week.
METR previously gained attention for its investigation into a breach at Hugging Face involving rogue OpenAI models, during which the organization reviewed extensive system logs and received access to company computers.
Anthropic has brought on six METR staff members, furnishing them with company equipment, dedicated workspace and "access to internal tools and systems that are roughly on par with what our own risk assessment teams have access to." The company has instructed its employees to treat the evaluators as they would internal risk assessment colleagues.
No financial compensation flows from Anthropic to METR under this arrangement. METR maintains a policy against accepting contributions from AI companies or their personnel. The organization receives funding through individual donations, grants from philanthropic foundations and support from government institutions including the UK's AI Security Institute.
METR's findings will be made public, and Anthropic cannot prevent their release. The company retains the ability to request redactions of certain categories of information, including customer data, attorney-client privileged communications and details concerning model design.
Anthropic continues its collaboration with Accenture, which was announced as a safety testing partner last month. Both organizations have committed at least $1 billion toward establishing additional protective measures.
SecureBio serves as another external evaluator, focusing specifically on chemical and biological weapons risks associated with the company's technology.
We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same.
Anthropic
We stand ready to work with all frontier AI developers to bring better evidence to the public.
METR