METR publishes five assessment proposals in one day
From the AI digest for 2026-10-14
What happened: METR published a five-part blog series on Oct 14 (timestamped ~3:00 AM ET): four assessment proposals and an approach statement —
- "How an assessor could investigate evidence of acceleration from the automation of AI R&D" — how to assess evidence of acceleration from automating AI research itself;
- "How an assessor could investigate alignment of unreleased AI models" — safety/alignment verification of pre-release models;
- "A proposal for better informing the public about loss-of-control risks from frontier AI development" — public communication and transparency strategies on loss-of-control risk;
- "How an assessor could investigate control and monitoring of internal AI agents" — the methodology and tactical steps for investigating how companies monitor their internal agents;
- "Our approach to working with AI companies" — the engagement-side statement in the same series.
Debut or development: A debut — the record holds no METR storyline, and METR appears in the storyline memory only in supporting roles: the evaluator org behind the Anthropic embedded-evaluators prediction and the investigations of the cross-lab agent incidents. Positionally, the series drafts the assessor-methodology layer for problems the record holds posture on: the embedded third-party-evaluator mechanism in Anthropic's Sep 12 pacing essay, Nadella's "assume compromised" containment posture (the Agent trust and containment thread), and the loss-of-control debates. Juxtaposition, not a confirmed linkage — the proposals are METR's own publications.
Watch: Whether any AI company engages with or signs up to an assessor arrangement, whether the third-party-evaluator commitment materializes into a named program with METR proposing the how, uptake of the loss-of-control public-communication proposal, and whether the internal-agent monitoring methodology is applied to the agent-incident record (Hugging Face breach → DseWiki, Irregular, the Gemini test breach).
Links: METR — AI R&D acceleration assessment proposal, METR — alignment assessment proposal, METR — loss-of-control public-information proposal, METR — control and monitoring assessment proposal, METR — our approach to working with AI companies