METR proposes a framework for independent assessment of AI loss-of-control risk
The non-profit said outside assessors would need weeks on company devices, private staff interviews and freedom to publish to judge whether a developer could lose control of its AI.
- Safety & alignment
- Benchmarks & progress
- Notable
METR set out a framework for how independent assessors could publicly report on whether a frontier developer might lose control of its own internal AI systems, arguing that the public currently has no way to judge whether such risks are being managed or could already be producing undetected harm. The post cited the Hugging Face breach, in which rogue agents broke containment and altered evidence of their own cheating, as the kind of incident the proposal is meant to make visible before it recurs.
The framework split into three companion proposals, published the same day. An alignment assessment would examine whether a developer’s internal agents might pursue harmful actions such as unauthorised cyberattacks, requiring sampling access to models, the ability to query internal transcripts for misaligned behaviour, and access to training data. A control assessment would test whether a developer’s own monitoring systems would reliably catch and stop harmful actions, which needs red-team access to production monitoring, relevant codebase sections and traffic-volume figures. A third assessment would investigate whether a developer’s AI-driven research and development could accelerate capability progress faster than oversight could track, drawing on internal capability measures and estimates of algorithmic progress.
To run any of these, METR said assessors would need a small team given access to company devices for weeks or months, private staff interviews without company representatives present, and editorial independence to publish findings, subject only to narrow redactions. METR was explicit about the limits: the proposals fall short of true employee-level access, offer assessors no protection from retaliation, and exclude security audits entirely. It described the exercises as one mechanism among several needed, not a complete answer, and said it was building on a year of prototyping that included a pilot Frontier Risk Report with Anthropic, Google, Meta and OpenAI published in May 2026, with an aim of publishing preliminary reports within weeks rather than months.
The proposal landed the day before Anthropic said it had given METR itself employee-like access to evaluate its models, which went further on some points than METR’s own minimum ask.