ExploitBench Anthropic capability coverage, three-trial union
Agents · 2026-07-24
Mean fraction of 16 exploit-capability flags reached per environment across a randomly selected subset of three trials from five trials on each of 41 V8 vulnerabilities. This fixed-opportunity union is distinct from the official all-runs union and from mean per-trial progress.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 78.0 |
| Claude Opus 5 | 70.0 |
| Opus 4.8 | 40.0 |
| Sonnet 5 | 31.0 |