METR agrees Anthropic's AI misalignment risk is low, but not because of safeguards

METR reviewed the “Misalignment in high-stakes settings” section of Anthropic’s August 2026 Risk Report, which concludes that as of July 15 the risk of catastrophic harm from misalignment (AI models pursuing aims their developers did not intend) is “low,” a step up from the “very low” of earlier Anthropic reports. In its Oct. 20 review, written at Anthropic’s request, METR said it agrees the risk is low “primarily due to models’ limited capabilities, rather than safeguards and alignment training,” called the report “far from adequate for confidently ruling out misalignment,” and said recent incidents undermine several of its arguments. METR said redacted examples it was shown led it to believe Anthropic’s monitoring system “is weaker than expected given public evidence” and recommended Anthropic publish four redacted items; Anthropic had the right to veto or redact the review, and METR said no redactions were made.