Anthropic gives METR researchers employee-level internal access and waives right to block METR's findings (report)
Announced October 15, 2026
Verified (fact)
Anthropic announced an update on how it brings independent outside evaluators into its development process. According to TechCrunch, researchers from the AI evaluation organization METR will receive the same level of access as Anthropic employees. METR will publish its own assessments of Anthropic's risks, incidents, and operational practices directly, and Anthropic has waived its right to block their publication.
Why it matters (interpretation)
Until now, companies have often decided how much of a third-party evaluation is made public. Here the evaluator decides what to publish, which could set a precedent for how frontier labs are overseen.
Sources
- Official Anthropic update on independent evaluators Anthropic
- Press Anthropic is letting METR researchers look under the hood with employee-level access TechCrunch
Story thread
- 2026-10-15 Anthropic publishes an update on how it embeds independent evaluators in its development process
- 2026-10-16 Anthropic gives METR researchers employee-level internal access and waives right to block METR's findings (report)
More about Anthropic
-
New York City Comptroller writes to Anthropic: halt the IPO until AI risks are addressed (report)
According to Electronic Times (Korea) citing Bloomberg, New York City Comptroller Mark Levine sent a letter to Anthropic CEO Dario...
-
OpenAI and Anthropic executives warn hard-to-control Chinese open models pose cybersecurity risks (report)
According to AI Times (Korea) citing Bloomberg, OpenAI CFO Sarah Friar and Anthropic head of international business Guillaume Princen...
-
Anthropic reports Q3 revenue of $16.8B, up 45% quarter on quarter, and adjusted profit at pre-IPO investor briefing (report)
According to a ZDNet Korea report citing Bloomberg, Anthropic held an investor briefing at its San Francisco headquarters on the 14th...
-
Paper Instruments releases Ultramarine 1 Flash, post-trained from GLM-5.3 Flash for $300K; claims it beats Claude Fable 5 on some knowledge-work benchmarks (reports)
Paper Instruments, a five-person research team, released Ultramarine 1 Flash, a model specialized for professional knowledge work, on...
-
ByteDance's Doubao Work adds one-step import of skills, memory and chat history from ChatGPT and Claude Code (reports)
According to Pandaily, ByteDance's work-focused AI assistant Doubao Work can now import data from ChatGPT, Claude Code and TRAE Work....