Anthropic Raises Its Own Misalignment Risk From 'Very Low' to 'Low' and Shelves an Internal Model 2
Unite.AI / Anthropic·high signal
The August 2026 risk report (covering February 24 through July 15) upgrades Anthropic's assessment of catastrophic harm from misalignment in high-stakes settings to 'low,' from 'very low' in the February report, and keeps non-novel weapons uplift at 'low but higher than our previous estimate' following the bioclassifier finding. The report also confirms an unreleased internal 'Model 2' that is somewhat more capable than Mythos 5 with no release plans, and that Fable 5 shares weights with Mythos 5 plus added safeguards. Multiple appendices — including bioclassifier exemption criteria and constitutional modifications — are redacted.