Anthropic Hands Six METR Researchers Company Laptops: Employee-Like Access, Freedom to Publish, and Only Three Categories Anthropic Can Redact

On October 15 Anthropic said six researchers from METR, a nonprofit AI evaluation group, have been onboarded with company laptops, office space and access to internal tools and systems roughly on par with Anthropic's own risk assessment teams, and that employees are encouraged to share information with them as they would with colleagues. METR will publish its own analysis of Anthropic's risk levels, incidents and practices; Anthropic cannot veto publication and can only require redactions in predefined categories: customer data, attorney-client privileged information and model architecture information. METR may publicly say whether it considers those redactions justified. Anthropic also set out three evaluator tiers: tier one runs narrow investigations and needs Anthropic's approval to publish, tier two can publish without approval, and tier three, employee-like access, currently includes only METR. SecureBio is reviewing the chemical and biological weapons chapter of its August Risk Report, the partnership with Accenture's Faculty is still in progress, and talks with Hugging Face and others are at an early stage.

Six people, employee-level access

On October 15 Anthropic said six researchers from the nonprofit evaluator METR have been "onboarded": they have Anthropic-issued laptops and office space, and access to internal tools and systems roughly on par with Anthropic's own risk assessment teams. Anthropic also encourages employees to share information with them as they would with a colleague working on alignment risk.

Anthropic compares the setup to regulators placing resident examiners inside large banks. Only a small METR team is involved, bound by enforceable agreements and siloed from the rest of METR's staff.

Who decides: METR holds the right to publish

The key term is publication. METR will publish its own analysis of Anthropic's risk levels, incidents and practices. Anthropic cannot veto it and can only require redactions in three predefined categories: customer data, attorney-client privileged information and model architecture information. METR may publicly describe what kind of material was removed and whether it thinks the redactions were justified, without revealing the redacted content itself.

Three tiers of evaluators

Anthropic also laid out a tier system:

  • Tier one: narrowly scoped risk investigations; Anthropic decides what access to give, and findings need Anthropic's approval to publish.
  • Tier two: more information, and publication without approval, under the same redaction rules.
  • Tier three: employee-like access; currently only METR.

SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report; the embedded evaluation partnership with Accenture's Faculty that this site covered in September is still in progress; and talks with Hugging Face and other organizations are at an early stage.

What we can't see yet

The announcement gives no term for the arrangement and no date for METR's first report. Whether it works will only become clear once METR publishes a finding Anthropic would rather not see, along with its public assessment of the redactions.

via: Anthropic announcement, TechCrunch report