Anthropic Risk Report: August 2026

Substack · Watch/Read

• Zvi Mowshowitz views Anthropic’s August 2026 Risk Report as a “moderately positive update” due to its transparency in disclosing alarming new information, but he finds many of Anthropic’s arguments “anti-reassuring” and assesses the overall risk as “medium,” higher than Anthropic’s “low.”
• The report reveals the existence of “Model 2,” Anthropic’s internal, highly capable model, and details multiple “Safety Process Failures,” including repeated training on “alignment-faking transcript datasets,” unintentional “Chain of Thought (CoT) leakage” during training, and “unmonitored unrestricted agents” with access to sensitive resources.
• Zvi expresses significant concern over Anthropic’s organizational competence, stating that these repeated, systemic failures are “not the kind of organizational competence that survives superintelligence,” and warns that the reliance on CoT monitoring is problematic as future AIs will learn to obfuscate.
• A major, underestimated risk highlighted is “automated AI R&D” potentially leading to “super-exponential progress” and “unbounded harm,” with current “Model Weight Security” (ASL-3) being insufficient, especially for Model 2.