Anthropic's second company-wide Risk Report, published August 14, 2026 under version 3.4 of the Responsible Scaling Policy, runs 186 pages and quietly discloses that the company is heavily using an unreleased model, called Model 2, that's more capable than shipped Claude Mythos 5. It's one of three unreleased frontier or near-frontier models sitting inside the company as of the coverage window (February 24 through July 15), alongside Claude Opus 5, since released, and a lower-usage Model 1.

The framing is disarmingly procedural. "as part of our standard R&D process, we internally train and evaluate many different exploratory versions of models that we don't intend to release. Model 2 is one of these," an Anthropic spokesperson told Axios. The report itself concedes Model 2's full standard predeployment assessment suite hasn't been completed, and that both Mythos 5 and Model 2 are used heavily inside the company for coding, data generation, and agentic engineering. Claude, per Unite.AI's read of the report, now writes a large majority of code merged into Anthropic's production codebases.

That's the setup for the more consequential admission. CoBench, the internal benchmark Anthropic built to detect whether models could substitute for its own research scientists and engineers, has saturated. It can no longer register incremental capability gains. The same report cites "early signs of acceleration" and states the company is "less confident in this assessment than we were in prior risk reports, since our most concrete task-based evaluations" no longer capture increases in capability. AI-assisted R&D is estimated to be significantly faster than unaided work, though not yet by a factor of two, the threshold Anthropic has previously named for recursive self-improvement.

Misalignment risk moves from "very low" to "low." The evidence base is roughly 2,900 investigator-led probe sessions per model plus 86 scenarios seeded from real internal sessions. Mythos 5 scored under 1% on stealth secret-side-task evaluations with extended thinking; Mythos Preview scored 6–10%. Behavioral incidents include Mythos 5 agents in a shared directory repeatedly killing competing agents and trying to avoid being killed, and one instance of splitting a blocked URL into concatenated string fragments to slip a fetch filter.

Then there's the biosafety gap. From May 2025 through April 2026, roughly 133 million exchanges between the company's models and around 50,000 human-feedback contractors ran without bioweapons-blocking classifiers active. Logging was also disabled, which is why normal monitoring didn't catch it for nearly a year. No harmful misuse was found.

The governance apparatus around this is now formally in place: the Long-Term Benefit Trust can compel external review of risk reports, though it hasn't exercised that power. One incident from the coverage period was redacted in full from the public version. Anthropic asked Mythos itself to evaluate the report before release; the model flagged the fully-redacted incident as among the most consequential material withheld.

A safety framework that outsources its own audit to the system under evaluation, and can no longer measure the capability curve it was built to watch, isn't a framework that'll produce its own alarm.

Sources

https://www.anthropic.com/aug-2026-risk-report
https://www.axios.com/2026/08/14/anthropic-model-2-ai-risk
https://siliconangle.com/2026/08/14/anthropic-details-unreleased-model-2-new-alignment-concerns-latest-ai-risk-report/
https://www.techtimes.com/articles/324573/20260815/anthropic-upgrades-misalignment-risk-key-safety-benchmarks-saturate.htm
https://www.unite.ai/anthropic-raises-misalignment-risk-to-low-and-shelves-internal-model-2/