The Pulse
METR Seeks Access to Frontier Labs for AI Risk Assessments
METR proposes independent assessments of how frontier AI labs manage misaligned agents, monitor internal systems and track AI-driven research acceleration. The voluntary process would let labs limit assessor access and redact reports, which
AI.info Team · Written with AI ·
METR wants frontier AI companies to let outside assessors examine internal agent behavior, but its proposal leaves the labs in control of access and redactions. The research nonprofit says public reports could expose evidence about loss-of-control risks; it also warns the process cannot establish that development is safe.
Published on October 14, the proposal asks companies to work with METR on assessments of AI systems used inside their own development processes. The assessments would focus on three questions: whether agents might attempt harmful actions, whether monitoring systems would catch them, and whether AI-driven research could accelerate capabilities faster than safeguards improve. METR says the public currently lacks enough evidence to judge how developers are addressing those risks.
Three questions behind the proposal
The alignment assessment would examine incidents observed inside a company and investigate whether its testing or training could hide more concerning behavior. A control assessment would probe how monitoring systems cover internal agents, whether those systems can be bypassed, and how staff respond to flagged activity. The third assessment would look for evidence of acceleration in internal capabilities, including whether AI agents contribute to that pace and what might constrain it.
METR says it has prototyped assessments of internal development with companies since 2025 and has worked on loss-of-control testing since 2022. Its May 2026 pilot Frontier Risk Report involved Anthropic, Google, Meta and OpenAI and included a stress test of internal monitoring systems. METR also says it conducted a brief exercise with OpenAI on the OpenAI–Hugging Face incident and has additional assessments underway, including an in-depth alignment assessment with Anthropic.
Labs would set the boundaries
Under the proposed terms, a small group of METR staff would receive access to relevant data on company devices for an agreed period, likely weeks or a few months. Assessors would interview relevant employees privately, without company representatives present or company screening of their answers. METR proposes writing reports independently within a scope agreed in advance, with companies allowed to redact sensitive information.
That arrangement is the proposal’s central tension: METR seeks independent scrutiny, but each company would choose what access to grant and could redact material. The nonprofit says reports should disclose the scope of an exercise, access limits and redactions, so readers can judge what the findings do—and do not—show. It proposes preliminary reports within weeks, while acknowledging that the short schedule may not produce firm conclusions.
METR says the work is not an audit
“The science of studying AI loss of control is in its infancy,” METR writes. The organization says its assessments would produce evidence about specific technical risks, not certify a lab as safe, enforce safety standards or decide whether a company should deploy a model. The work would also leave out areas such as model-weight security, sandbox security and the quality of a company’s internal governance.
METR acknowledges that voluntary engagements can create pressure on assessors to avoid alienating the companies they depend on for access, and that companies can end an engagement. Its proposal also does not provide the employee-level access described in a recent public letter or guarantee protection from retaliation over unfavorable findings. The nonprofit says developers can contact it to arrange assessments and that it plans to share more about which companies take part.