Loading post…
AnthropicAnthropicAnnouncement
An update on independent evaluators
Summary
Anthropic has begun providing employee-like access to METR, a nonprofit evaluator of frontier AI models, as part of its commitment to independent third-party assessment. The company signed the White House Accord on Super Intelligence and aims to close the gap between what evaluators and the company itself understand about its systems. To evaluate frontier models, evaluators require expertise in capabilities assessment, independence, deep access to training and safeguards, and freedom to publish results. Anthropic defined three tiers of access: the first tier allows publication only with company approval, the second tier permits publication without approval but with possible redactions, and the third tier grants employee-like access with similar publication rights as the second. METR is currently at the third tier, while SecureBio operates at the first tier, and collaborations with Accenture and Faculty are in early stages.
Context
The announcement comes amid growing calls for independent oversight of frontier AI models, following the White House Accord on Super Intelligence. Anthropic aims to set a precedent for voluntary evaluation partnerships to improve safety transparency while protecting sensitive information.
Approach / What changed
Anthropic implemented a three-tier evaluator access system, granting METR employee-like access to internal tools and systems, and onboarding SecureBio at the first tier and collaborating with Accenture and Faculty with ambitions for deeper integration.
Takeaways
METR receives employee-like access including company hardware, office space, and internal tools, enabling deep evaluation without ability to veto their published results except for narrow categories like customer data or model architecture.
Anthropic established three tiers of evaluator access: first tier requires company approval for publication; second tier permits publication without approval but allows redaction of specific information categories; third tier (METR) has similar publication rights to the second.
Additional evaluators like SecureBio are at the first tier, while partnerships with Accenture and Faculty are in early stages with ambitious plans for model evaluation, red-teaming, and alignment assessment.
FAQ
What does 'employee-like access' mean for METR?
Anthropic onboarded six METR employees with company hardware, office space, and access to internal tools comparable to its own risk assessment teams. They can access information except customer data, attorney-client-privileged info, or model architecture details, and can publish findings without Anthropic veto except those categories.
Why does Anthropic use independent evaluators instead of public disclosure?
Publicly sharing all safety-relevant information could aid malicious actors or reveal commercial secrets. Instead, they provide deep access to third-party evaluators who publish evidence-based conclusions about risks, aiming to close the gap between what evaluators and the company understand.
What are the three tiers of evaluator access?
First tier: evaluators conduct scoped investigations and publish only with Anthropic approval. Second tier: evaluators can publish without approval; Anthropic can require redaction of specific categories. Third tier (METR): employee-like access with publication rights similar to second tier.
Source and freshness
- Source
- anthropic.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.