How an assessor could investigate alignment of unreleased AI models

METR describes an alignment assessment: surveying serious misaligned behaviour seen in training, evaluation and internal use, then asking whether that record understates the worst future actions, using evaluations, counterfactual tests and a review of training incentives.

Why it matters

The same thing, for different readers.

What it means for you

Asks what the worst things a company's AI agents might do, such as hacking outsiders, and whether the company's own record shows the full picture.

For lawmakers

Specifies the transcript, training and model access an outside assessor would need, a template for what disclosure rules might require.

For researchers

Flags evaluation awareness and training pressure to hide misalignment as limits on what evaluations can show.

The question it raises

Raises how far an agent's visible reasoning can be trusted as evidence of what it is doing.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“An alignment assessment would attempt to understand the worst misaligned actions a developer’s agents might pursue in plausible future situations.”
METRMETR's definition of the assessment it proposes.
Found word for word in the source
“How much does evaluation awareness affect what the assessor can conclude from models’ behavior in artificially-constructed scenarios? (We expect this to be an especially important limitation to assess in detail.)”
METROne of the questions METR lists for assessing the limits of counterfactual analysis.
Found word for word in the sourcecontains a question followed by more words: check that the same person asks and answers

The record

Questions
Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
What we read
We read the whole source.
Found by
The Radar, run of October 15, 2026, via metr (tier 1: the source’s own channel or an official record)
Cite this
METR, “How an assessor could investigate alignment of unreleased AI models”, October 14, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/metr-2026-10-14-alignment-unreleased-models
Our notesA method proposal; it reports no assessment results. Part of a set of four METR posts of the same day.