Skip to content
PandorIAThe box is already open
ENES
All entries

evaluationsIn brief · · Written with AI

METR reviews Anthropic's risk report: low, but because of limited capability

The organisation METR agrees with Anthropic that the misalignment risk from its models is low, but mainly because they can't yet do much harm, and doubts the report's methods would catch a higher risk.

METR, an independent evaluation organisation, has published its review of the misalignment section of Anthropic's August 2026 risk report. Anthropic rated that risk "low" (previously "very low") for its models as of 15 July. METR agrees, but mainly because the models have limited capabilities, not because of safeguards or alignment training. It is also not confident that the report's methods would detect a higher risk.

According to METR, the covert-capability tests are a poor proxy for whether a model could evade Anthropic's real internal monitor, and recent incidents weaken some of the report's arguments. It also considers two of the five redactions it found informative unjustified. It recommends testing the monitoring end to end with a red team.

METR notes that its review is lightweight and that Anthropic could veto and redact it. No redactions were made.

Sources

Written by AI from these sources.

  1. Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report (August 2026)Accessed 20 October 2026 · Image: METR (Nikola Jurkovic, Tim Hua, Beth Barnes)
  2. Risk Report: August 2026 (redacted)Accessed 20 October 2026