Anthropic Mythos 5.2 ExploitBench - Plain Cap Percent: leaderboard

Anthropic's final Claude Mythos 5.2 evaluation on 41 V8 vulnerabilities, with five plain-arm trials per vulnerability and a 300-turn budget in the authors' static uniform harness. Cap% measures the available exploitation capability flags captured across a random three of the five trials in each environment, averaged across environments. It is a percentage of captured flags, not ordinary attempt accuracy. Cyber safeguards are disabled; sandbox escape classifiers remain enabled. Reasoning effort is unreported and the plain-arm published roster contains only Mythos 5.2.

Metric: Flags captured across three of five trials, per-environment mean (%). Source: assets.claude.com. 1 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=anthropic-mythos-5-2-exploitbench-plain-cap-percent · How It Works · Data refreshed daily, snapshot 2026-11-12.