Anthropic reports 89.6% for Claude Mythos 5.2 on CVE Rediscovery, Anthropic's internal cybersecurity evaluation. The evaluation is Anthropic's own internal test of finding published vulnerabilities from pre-fix code snapshots.
Tasks: Published CVEs found from a project snapshot taken before each fix, as Anthropic describes it.
Setting: Highest-available effort, per the table footnote; the effort level is not named for this row. Numeric reasoning and token budgets are not stated.
Limits: Provider-reported by Anthropic in its launch announcement; not independently reproduced here. Anthropic points to the Fable 5.2 and Mythos 5.2 system card for methods; that card was not opened, so task subset, sample size, harness, prompts, trial counts and run dates are unverified. Results under the same benchmark name can be shown together as reported scores, but unmatched settings and missing methods rule out a controlled ranking. Internal Anthropic evaluation, not a public benchmark with an independent owner. Anthropic mentions an October 2026 run against 1,829 CVEs, but does not tie the table figure to that run explicitly; sample size is left null.
Anthropic
