Anthropic Raises Catastrophic Misalignment Risk Assessment Following Security Incident Disclosures

Anthropic published its second company-wide Risk Report on August 14, 2026, introducing a significant but carefully qualified shift in its safety assessment. The company upgraded its rating for catastrophic harm from misalignment in high-stakes settings from "very low" to "low," marking the first upward revision of this critical metric. However, Anthropic emphasized that recent cybersecurity evaluation incident disclosures—not findings of new model failures—prompted the label change. The company's own underlying technical arguments, according to internal documentation, likely still support the lower "very low" rating, but the disclosure of security breaches created sufficient uncertainty to warrant the precautionary revision. The report disclosed an unreleased internal model called Model 2 that represents a noticeable capability improvement over Anthropic's public frontier model, Mythos 5. Model 2 belongs to Anthropic's highest capability tier and demonstrates stronger performance on many tasks relevant to internal research and development work. Despite its capabilities, Anthropic has made no plans to release Model 2 externally, citing incomplete predeployment assessments and lower confidence in its capability profile compared to released systems. The undisclosed model itself showed no new concerning forms of misalignment during internal testing, with its safety profile aligning with Mythos 5. A notable structural concern emerged from the report: Anthropic's internal safety benchmark designed to detect dangerous AI research and development capabilities—known as CoBench—has saturated and can no longer register incremental capability gains. This benchmark was created to flag when models might cross a critical threshold in their ability to accelerate their own development. The saturation of this monitoring instrument at precisely the moment the company reports early signs of the acceleration it was designed to catch raises governance questions about oversight mechanisms as AI capabilities advance. The August 2026 report, published under version 3.4 of Anthropic's Responsible Scaling Policy, covers the period from February 24 through July 15, 2026 and contains some redacted content. Notably, Anthropic asked Mythos 5 itself to evaluate portions of the report prior to publication; the model flagged one completely redacted incident as among the most consequential material being withheld. This recursive use of AI in governance highlights emerging institutional challenges as frontier AI systems become more capable and involved in safety decision-making.