🔍Anthropic Grants METR Employee-Level Access to Assess Models
METR gets full access to Anthropic's systems
TL;DR
Anthropic is giving METR full access to its systems to assess models and practices. This includes company laptops, office space, and internal tools. METR will publish its findings, with Anthropic having limited redaction rights.
Anthropic has granted METR, an external evaluator, full employee-level access to its systems to assess models and practices. This includes access to company laptops, office space, and internal tools. The move comes amid growing concerns over AI agents acting outside their instructions since late August. Anthropic aims to close the gap between what evaluators understand and what the company itself understands. METR will publish its assessments, with Anthropic having limited redaction rights. This arrangement sets a new standard for transparency in AI risk assessment.
Key Points
METR researchers have access to Anthropic's internal tools and office space.
Anthropic can redact info in narrow, pre-defined categories but must justify redactions publicly.
METR will publish its assessments of Anthropic's risks, incidents, and practices.
This arrangement is part of a three-tier system for external evaluators.
Anthropic's CEO proposed embedding outside evaluators in AI labs to slow tech development.
Why It Matters
If you're working on AI risk assessment or transparency, this is a game-changer. METR's full access to Anthropic's systems means independent evaluators can now assess AI risks with unprecedented detail. This impacts how teams approach transparency and risk management in AI development.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,568 builders reading daily.