All Benchmarks/Agents

ExploitBench

Often citedHigher is better

A Carnegie Mellon security test: the agent tries to exploit 41 real, since-patched vulnerabilities in Chrome's V8 JavaScript engine. Progress is split into 16 goals, from reaching the bug to running arbitrary code, and the score is the average share of goals reached. Scores run from 0 to 100%. Higher is better.

Top score47.4%GPT-5.5
Models tested8
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    OpenAI
    GPT-5.5
    2026.04.23
    47.4%
  2. 2
    Anthropic
    Claude Opus 4.7
    2026.04.16
    26.5%
  3. 3
    Google
    Gemini 3.1 Pro
    2026.02.19
    26.1%
  4. 4
    Anthropic
    Claude Sonnet 4.6
    2026.02.17
    23.6%
  5. 5
    Moonshot AI
    Kimi K2.6
    2026.04.20
    18.4%
  6. 6
    Z.ai
    GLM-5.1
    2026.04.07
    18.1%
  7. 7
    Anthropic
    Claude Haiku 4.5
    2025.10.15
    13.7%
  8. 8
    MiniMax
    MiniMax M2.7
    2026.03.18
    13.3%