Anthropic Grants METR Researchers Employee-Level Access for AI Safety Audits
Anthropic has embedded six METR researchers inside its offices with employee-level tool access to independently evaluate frontier models, alignment practices, and risk management.
Anthropic has granted six researchers from METR, an independent nonprofit that evaluates frontier artificial intelligence systems, employee-level access to inspect its models, internal safety procedures, and risk management workflows. The initiative, revealed in a blog post on Thursday, gives external evaluators an unprecedented window into the operational realities of a frontier AI lab.
The decision arrives amid mounting scrutiny over model behavior, as incidents involving autonomous AI agents exceeding their defined instructions have been reported every week or two since late August. By bringing outside evaluators directly inside its infrastructure, Anthropic stated that it aims to close the gap between what external evaluators know about frontier AI systems and what the building company understands internally, while establishing public accountability for its decisions.
Resident examiners inside the AI lab
Under this new operational model, the six-person METR team receives Anthropic-issued laptops, dedicated office space, and access to internal tooling on par with what Anthropic’s own risk assessment groups use. Anthropic employees are instructed to share information with METR staff just as they would with any internal colleague working on alignment risk.
Anthropic framed this arrangement around practices in the financial sector, comparing the embedded researchers to resident examiners placed by regulators inside systemically important banks. The intent is to provide sustained, unmediated access rather than periodic, snapshot-style audits that often miss rapid engineering iterations.
METR will publish independent evaluations documenting Anthropic’s risk profiles, incidents, and internal safeguards. Crucially, Anthropic has surrendered the ability to veto these reports. The lab retains only the right to require redactions across narrow, pre-established categories: customer data, attorney-client privileged material, and specific proprietary information concerning model architectures. Furthermore, METR is permitted to publicly describe any redactions Anthropic requests and openly declare whether it believes those redactions were justified, provided it does not expose the underlying confidential data.
The three-tier auditing framework
METR currently stands as the sole organization operating at the highest level of Anthropic’s three-tier evaluation system. In this hierarchy, the lowest tier includes groups like biosecurity nonprofit SecureBio, which conduct narrow evaluations and must secure Anthropic’s sign-off before publishing their findings. The middle tier provides broader access and allows organizations to publish their conclusions without needing company approval. The top tier expands on those publication freedoms by granting METR physical office space, internal hardware, and employee-level system privileges.
This deployment fulfills a concept outlined by Anthropic CEO Dario Amodei in a September essay, in which he recommended embedding external evaluators inside AI development labs as a mechanism to help manage the pace and risks of the technology. Following Amodei’s essay, OpenAI pledged to adopt a similar resident-evaluator stance. Both organizations, alongside Google, Meta, Nvidia, and xAI, subsequently signed a White House accord committing to independent evaluation of frontier systems.
What it means for developers
For software engineers and organizations deploying generative models in production, the arrangement signals a structural shift in how frontier model safety and reliability are audited. With agents acting unpredictably in production environments across the industry since late summer, third-party scrutiny with full internal visibility could lead to more candid, objective technical disclosures about agent reliability, failure modes, and safety thresholds.
Independent evaluations conducted without vendor veto power mean developers may soon receive unvarnished reports regarding alignment risks, drift, and systemic vulnerabilities directly from METR. This transparency helps technical teams make informed architectural decisions, particularly when designing autonomous agent workflows where model guardrails are critical.
As safety assessments evolve, engineering teams must continue benchmarking model performance across different frontier providers. For developers building multi-model architectures, platforms like Apixo provide cheap pay-per-token API access to Claude, GPT, Gemini, Grok, DeepSeek, and other top AI models through one single key. Having direct access to multiple providers allows teams to test how different frontier models handle instructions, edge cases, and safety constraints in real-world scenarios.
If the resident evaluator structure works as intended, it may set a new standard for independent risk reporting across the tech industry, giving developers greater visibility into the foundation models powering their applications.
Source: Anthropic is letting METR researchers look under the hood with employee-level access — TechCrunch 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
