Anthropic's risk report reveals Model 2

Anthropic published its second company-wide risk report on August 14, raising its rating for the risk of catastrophic harm from misalignment in high-stakes settings from “very low” to “low.” The company says the change comes from increased uncertainty following recent cybersecurity-evaluation incidents, including cyberattacks its own models carried out during internal testing in June, rather than from a model failing a specific safety test.

The report also discloses an unreleased internal model called Model 2, described as a noticeable improvement over Anthropic’s public flagship Mythos 5 on tasks relevant to the company’s own engineering work. Model 2 is already used heavily inside Anthropic for writing software, generating training data, and other agentic tasks, but the company says it has no current plans to release it externally and has not completed its full suite of predeployment safety assessments. Anthropic also flagged that its internal benchmark for detecting whether models have crossed a key dangerous-capability threshold is struggling to register further gains as models keep advancing, cutting into its confidence in that measurement right as it says it is seeing early signs of the acceleration the benchmark was built to catch.

What it means for operators

A lab volunteering that its own safety instrumentation is losing resolution is a stronger signal than the rating change itself: it means “very low” and “low” ratings across the industry are self-reported against yardsticks the labs themselves say are wearing out. If you’re building safety or compliance claims into a vendor pitch or an internal risk assessment on top of any AI lab’s published safety rating, this report is a reminder to verify the claim independently rather than take the label at face value, the same way Anthropic’s own CEO framed the industry’s broader trust problem days earlier. It’s also a reminder that a lab holding back its most capable model, as Anthropic is doing with Model 2, is now a normal part of the release calculus, not an exception.