Anthropic details METR evaluator access, with confidentiality and publication limits

Read the original article →

Anthropic said on October 15, 2026 that it had begun giving six METR employees access to internal tools, systems, hardware and office space broadly comparable to its risk-assessment teams. Its announcement was clarified on October 16.

What the arrangement permits

According to Anthropic, METR can publish assessments without a company veto. However, Anthropic can require redactions under predefined confidentiality categories and ultimately decides whether information falls within those categories. METR retains the right to make certain statements about the access it received.

The access has exceptions, including customer data, privileged legal information and model-architecture information. The embedded team is separated from other METR staff. Anthropic says other evaluator partnerships are at earlier stages, with more limited access and publication rights. This is a voluntary arrangement, not a published safety certification.

MintedBrain analysis

For readers comparing AI safety claims, the useful questions are concrete: What could the evaluator inspect? What could it publish? Which limits remain? An access announcement is a starting point for scrutiny. Readers still need the resulting assessments and their scope before drawing conclusions about a model.

Explore MintedBrain Academy for AI learning resources.

Source: Anthropic announcement, October 15, with October 16 clarification. Reviewed November 3, 2026. No model testing was performed for this summary.

References

This article was originally published at Anthropic. For the full piece, read the original article.

Discussion

  • Loading…

← Back to News