METR proposes independent AI risk reviews, with labs controlling access and redactions

METR wants outside assessors to examine what happens inside AI labs, not just how their finished models perform. In a proposal published October 14, 2026, it outlined public assessments of whether advanced AI agents might evade human control, whether safeguards would stop them, and whether AI-assisted research could outrun developers’ ability to manage the risks.

The proposed work focuses on AI used internally during development, rather than only on systems released to customers. METR argues that the public lacks enough visibility to judge whether developers are addressing concerning agent behavior—or whether more serious failures could follow as models become more capable.

This builds on assessment exercises METR says it has been developing with companies since 2025. Its May pilot Frontier Risk Report involved Anthropic, Google, Meta and OpenAI and included a stress test of internal monitoring. The new proposal describes how developers could give outside investigators access to evidence for public reports.

Three questions, from agent behavior to research speed

• Alignment: Do agents act as their developers and users intend? Assessors would examine internally observed incidents, how companies find them, and whether those methods miss more serious behavior. They would also investigate whether training hides signs of unwanted behavior rather than eliminating it.
• Control: Would safeguards catch and stop an agent attempting harmful actions? The assessment would examine monitoring coverage, resistance to attacks by agents, and reliability in flagging concerning behavior. It would also test how quickly and reliably the company responds, and what could happen if monitoring fails.
• AI research acceleration: Is AI speeding up the development of more capable AI? Assessors would examine internal capability gains and progress in the methods used to build models. They would ask whether AI caused any acceleration and whether limits such as computing capacity or data might constrain further gains.

The research-speed question requires a distinction between producing more work and improving AI faster. METR says increases in code output or experiments do not by themselves reveal how quickly model capabilities are advancing. Its proposed assessment would look for that connection, rather than treating a busier development process as proof of faster progress.

We instead aim to get new independent evidence out to the public so that company staff, boards, external scientists, and policymakers are able to arrive at their own conclusions.

METR, October 14 proposal

Private interviews, negotiated access

METR would write reports using staff interviews and internal datasets and systems accessed on company devices. A small team would work within a scope agreed in advance, likely for weeks or a few months, with help from a technical contact at the developer.

Relevant employees would speak privately, without a company representative present or management screening their answers. METR also asks that companies not monitor those conversations, so staff can raise concerns candidly. The terms call for timely publication of METR-written reports within the agreed scope, subject to redaction.

Developers could choose what access to grant METR and redact information they consider sensitive. In return, METR asks for transparency about the assessment’s scope, access, redactions and other key limitations. Those disclosures would accompany the findings, showing readers the boundaries of the investigation.

Evidence, not a safety certificate

METR explicitly says the proposal cannot confidently establish that development is safe. It says the science is not mature enough for consensus conclusions, and there are no safe-development standards against which to audit compliance. The reports are intended to inform judgments, not certify that a developer has passed a settled safety test.

The proposal also stops short of true employee-level access and cannot fully protect against retaliation over adverse findings. It excludes security audits and is not a substitute for other transparency efforts or enforcement of safety standards. METR is inviting frontier developers to participate and says it plans to share more about which assessments companies undertake.