METR proposes how an outside assessor could check AI alignment before release

In short

METR, a research nonprofit that tests advanced AI systems, has published a proposal for how an outside assessor could check an AI developer's unreleased models for serious misaligned behavior. "Misaligned" here means acting against what its developers intend. The report is a plan. It does not describe results from a completed assessment.

What happened

METR published the proposal on October 14, 2026. It says developers and the public want to know whether AI agents would take very harmful actions, such as unsanctioned cyberattacks. The assessment would try to understand the worst misaligned actions a developer's agents might take in plausible future situations.

METR describes two main parts.

• A survey of observed incidents. The assessor would review serious misaligned behavior seen in training, evaluations and internal use across all near-frontier internal models. Near-frontier means close to the most capable models the developer has. The review would count incidents and study the most severe ones in detail.
• A check on whether those incidents tell the full story. The assessor would ask whether the behavior seen understates what agents might do in other situations. This would use evaluations and counterfactual simulations, which test how agents might act if circumstances were different. It would also ask whether training might push agents to hide problems, or be tuned to the assessor's own detection methods.

The behaviors in scope include hacking an outside party, tampering with logs, deceiving human operators, colluding to undermine oversight, setting up unsanctioned deployments, and acquiring resources beyond a task's scope.

METR names limits. It says evaluation awareness, where a model may notice it is being tested, will likely limit how far the assessor can rely on simulations. It also asks the assessor to report whether logs are complete and whether models' written reasoning can be trusted.

On access, METR says the assessor would need sampling access to internal model checkpoints and the ability to query a broad set of internal transcripts, including from training. Staff interviews would also be needed. METR says that during its Hugging Face investigation it spent roughly $400,000 in API credits over six days. It estimates a rigorous two-month investigation of this scope would cost over $10 million.

What was done about it

This is a proposal, so no response is described. METR suggests ways to handle developer concerns. Developers could filter out very sensitive transcripts if they explain how the filters work. If they cannot give direct access, an internal researcher could run queries for the assessor, though this would take more staff time.

Why it matters

METR says the aim is to understand the worst misaligned actions a developer's agents might take in plausible future situations, not only those already seen. It notes that observed incidents may not be representative. The report does not say any developer has adopted the plan. Read the full details in the original report.