METR agrees Anthropic's misalignment risk is low, but says that is mainly because the models are not yet capable enough, not because of safeguards

METR, the AI evaluation group, published on 20 October its review of the “Misalignment in high-stakes settings” section of Anthropic's August 2026 Risk Report. The authors are Nikola Jurkovic, Tim Hua and Beth Barnes.

METR agrees that the risk is low, but says this is mainly because of the models' limited capabilities rather than safeguards and alignment training. It writes: “We are not confident that, if risk were higher, the methodology in the report would reliably identify this.” It says the benchmarks Anthropic used for hidden capabilities are a poor stand-in for evading Anthropic's own monitoring, and that recent incidents weaken several of the report's arguments.

METR recommends end-to-end red-teaming of Anthropic's monitoring systems, and adding to the report the risk of models undermining safety decisions inside Anthropic. It asks Anthropic to publish four pieces of information that were redacted.

On access, METR says it saw unredacted versions of three sections of the report. Anthropic had the right to veto or redact the review; METR says nothing was redacted and its conclusions did not change, though it made minor wording changes after Anthropic's feedback. METR says its authors were engaged before Anthropic's announcement that METR could publish without its approval, which the Switchboard covered on 16 October.

• Confirmed METR agrees the misalignment risk is low but says this is mainly due to the models' limited capabilities, not safeguards and alignment training. METR
• Confirmed Anthropic could veto or redact the review; METR says no redactions were made. METR