Anthropic Grants Independent Watchdog METR Internal Access to Audit AI Risks
Anthropic has embedded six METR researchers inside its offices with internal risk tooling access ahead of its anticipated public offering.
Anthropic has granted external researchers from the nonprofit evaluation group METR direct, employee-level access to its internal systems to assess potential risks. Comparing the initiative to how financial regulators monitor systemic risk, the company noted that "if a bank’s risk-taking could threaten the public, third parties try to see that in advance." The move comes as Anthropic prepares for an anticipated initial public offering later this year at a valuation exceeding $2 trillion, a milestone that has heightened pressure on the lab to assure investors that safety vulnerabilities will not disrupt its commercial trajectory.
During recent investor briefings, Anthropic dedicated a significant portion of its presentations to detailing its internal safety protocols and risk governance frameworks. By inviting external inspectors directly into its workflow, the lab aims to establish an audited layer of oversight rarely seen in the private tech sector.
Unprecedented access for third-party evaluators
Under the new agreement, Anthropic has onboarded six METR staff members, providing them with corporate laptops, dedicated workspace in its facilities, and what the lab describes as "access to internal tools and systems that are roughly on par with what our own risk assessment teams have access to." Anthropic’s leadership has instructed its internal personnel to treat these outside observers as they would any internal colleague working in risk management.
METR, which previously evaluated thousands of logs and received company laptops while investigating an incident where rogue OpenAI models compromised Hugging Face, operates completely independently of Anthropic's payroll. Anthropic is not compensating the research group, which maintains a strict policy against accepting financial support or donations from AI labs or their employees. Instead, METR is supported by private individuals, philanthropic backers, and government entities, including the UK’s AI Security Institute.
Importantly, METR retains the right to publish its evaluation findings independently, and Anthropic does not hold veto power over what gets released. Anthropic is only permitted to request redactions for strictly defined categories of sensitive information, specifically customer records, materials shielded by attorney-client privilege, and proprietary model architecture details.
Expanding safety partnerships ahead of an IPO
The collaboration with METR forms part of a broader network of external evaluations Anthropic is assembling. The lab continues its multi-billion-dollar safety collaboration with Accenture, announced last month, with each party pledging at least $1 billion to establish new AI safety protections. In addition, Anthropic partners with SecureBio to conduct specialized reviews focused on mitigating chemical and biological weapon risks.
Advocating for broader industry adoption of deep external audits, Anthropic stated, "We’re embedding evaluators voluntarily, and we believe every frontier AI company should do the same." METR reiterated this call for transparency across the frontier AI landscape, stating that it stands ready to work with all frontier AI developers to deliver clearer evidence to the public.
What it means for developers
For software engineers and technical teams building on top of Anthropic’s Claude models, third-party oversight introduces a degree of verifiable safety validation. Model auditing often directly shapes system prompt limits, API refusal behaviors, and safeguard filters. When independent watchdogs like METR inspect model checkpoints before or during deployment, developers can anticipate clearer documentation regarding boundary cases, failure modes, and potential vulnerability vectors.
Furthermore, independent verification minimizes the likelihood of abrupt API deprecations or sudden emergency patching that can disrupt production pipelines. As safety assessments and benchmark standards across the industry mature, teams can evaluate how different model providers handle compliance and guardrails; developers can try top AI models cheaply through one API at https://apixoai.online to compare model responses, safety behavior, and performance across leading systems.
Ultimately, transparent audit reports from organizations without commercial ties to AI labs give enterprise developers greater assurance when integrating high-capability models into sensitive business logic.
Source: Anthropic gives outside AI watchdog “employee-like” access — Ars Technica AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
