METR published a proposal on Oct. 14, 2026, for more public evidence about risks from AI developers’ internal systems, alongside an explanation of its approach to working with those companies. The documents give business readers a concrete way to distinguish external assessment from a broad assurance that development is safe.
Rapid News Brief
A proposal for evidence, not certification
In its proposal, METR identifies three areas for assessment: whether agents might act against intended constraints, whether monitoring would catch harmful behavior and whether AI-assisted research could accelerate development beyond existing controls.
METR proposes reports informed by staff interviews and access to internal data and systems. Developers would choose the access provided and could redact sensitive information, while the resulting reports would describe important limitations. The organization explicitly says the proposal would not provide confident assurance that development is safe and is not a substitute for other safety efforts.
The assessment process is part of the result
In a companion explanation, METR says readers should know the scope and duration of an engagement, access restrictions, redactions and other pressures that could affect conclusions. It also describes policies intended to disclose conflicts and maintain financial and operational independence.
METR says voluntary cooperation remains a limitation: companies decide whether to engage, provide access and permit publication. It describes evidence collection as its role, rather than deciding how developers should deploy their agents. These are METR’s stated policies and proposed terms, not proof that every assessment has unrestricted access.
What business readers should inspect
The practical distinction is between an assessor’s participation and the evidence a particular report actually supplies. A reader can inspect which systems were examined, what was excluded and whether a conclusion applies to the use being considered.
A limited assessment can still provide useful evidence, but its limitations should travel with any executive summary. Neither the presence of an external evaluator nor a report’s title should be treated as a universal safety guarantee for unrelated products or workflows.