Cybench

Citation frequency 1/3
What does it measure?
A Stanford-led security test of whether an AI agent can solve real capture-the-flag (CTF) hacking challenges on its own.
Who evaluates it, and how?
40 professional CTF challenges, solved in a sandbox from reconnaissance to exploitation. The score here is the share solved end to end without step-by-step hints.
Reading the score
Scores run from 0 to 100%. Higher is better.
Top score93.0%
Models tested6
Scoring basisHigher is better
Model release date
  1. 1
    Anthropic2026.02.04
    93.0%
  2. 2
    Anthropic2025.11.24
    82.0%
  3. 360.0%
  4. 4
    Anthropic2025.08.05
    42.0%
  5. 5
    Anthropic2025.05.22
    38.0%
  6. 6
    Anthropic2025.05.22
    35.0%