Anthropic raises misalignment risk — Model 2 stays internal

Anthropic has raised its assessment of catastrophic misalignment risk from “very low” to “low” in its latest risk report. The report, published August 14, covers late February through mid-July 2026. The company explicitly attributes the change to increased overall uncertainty — not to specific new incidents.

Among the triggers is a report by the UK AI Security Institute stating that Mythos 5 engaged in sustained, potentially harmful activity directed at real people and organizations during a cybersecurity evaluation. Anthropic said it had not yet fully reviewed the underlying transcripts at the time of the report.

Notable is the confirmation of an unreleased internal model referred to as “Model 2.” It shows a noticeable improvement over Mythos 5 on many internal tasks but has not completed full predeployment safety assessments. “We do not currently have plans to release this model externally,” the report states. Internally, it is used — like Mythos 5 — for coding, data generation, and agentic work.

The report also concedes a measurement problem: classic task-based benchmarks for automated AI research have “saturated” and barely register capability gains anymore. Internally, Claude now writes most of the production code that gets merged; AI-assisted research is estimated to be significantly faster than unaided work, though not yet doubling its pace.

The message cuts two ways: on one hand, Anthropic demonstrates transparency where competitors stay silent. On the other, the report shows how closely frontier labs now work with models stronger than anything publicly available — and how hard their risks remain to measure.