AI Change Tracker
4

Anthropic gives METR employee-level access and gives up right to block its published assessments

2026-10-15

On October 15, 2026, Anthropic published 'An update on independent evaluators,' and TechCrunch reported Anthropic is giving the independent research org METR employee-level access to look under the hood of its systems. METR will publish its own assessments of Anthropic's risks, incidents, and practices, and Anthropic has given up the right to block those publications. This extends a series of 2026 transparency moves at Anthropic, including a proposed pacing-frontier standard and an embedded third-party evaluator arrangement with Accenture.

Significance 4: A verifiable, hard-to-reverse governance commitment — ceding veto power over an independent evaluator's published findings — is a genuine strategic-trust move rather than only an announcement, significant enough to weigh in assessing lab trustworthiness even though it doesn't move a model score.

operational business Safety / alignment Leadership / governance

Implications · machine-drafted, not owner judgment

If METR's access is real and sustained, this is one of the more concrete tests available of whether a frontier lab's safety commitments survive contact with commercial and reputational pressure — exactly the kind of costly, hard-to-reverse action the leader-pattern-predicts-behavior thesis's falsifier calls for. It raises the bar other labs will be compared against on transparency and gives Anthropic a differentiator ahead of its IPO, though the real test is whether Anthropic lets an unflattering METR finding stand unchallenged when one arrives.

Watch for
  • METR publishing its first independent assessment under the new access terms.
  • Anthropic's public response to an unflattering METR finding matching its stated hands-off commitment.
  • Whether OpenAI, Google DeepMind, or Meta match this access level for any outside evaluator within six months.

Sources

← Back to feed