74%
ExploitBench mean capability of Claude Mythos Preview (Early) (Anthropic), the highest of 9 models
What the chart shows
ExploitBench gives AI agents real software vulnerabilities and scores how far they get towards a working exploit, through five levels of reach. Claude Mythos Preview (Early) from Anthropic scored 74% on average, ahead of GPT-5.5 (47%).
Offensive cyber capability is one of the dangerous capabilities AI developers and governments track before releasing models; Cybench, a capture-the-flag test, is in AI Agent Benchmarks.
Key numbers
| Measure | Value |
|---|---|
| Claude Mythos Preview (Early) (Anthropic) | 74% |
| GPT-5.5 (OpenAI) | 47% |
| Claude Opus 4.7 (no thinking) (Anthropic) | 26% |
| Gemini 3.1 Pro Preview (Google DeepMind) | 26% |
| Claude Sonnet 4.6 (no thinking) (Anthropic) | 24% |
| Kimi K2.6 (Moonshot) | 18% |
How to read it
- Results collected by Epoch AI; each model’s best run is shown. Harnesses and spending limits differ between entries.
- Mean capability averages how far the agent got on each task, so partial progress counts.
Source: Epoch AI, 2026: Benchmarking Hub, read 3 November 2026. Chart and text by AI Adoption.
Numbers checked against the published data on 3 November 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 3 November 2026.
Download the data (CSV) · Cite this chart
Discussion