ExploitBench
Exploit development across tiers of difficulty.
Agent Index weight
0%
shown for context only
Models tested
9
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | Claude Mythos Preview Anthropic | 73.8% | source | — | — | |
| 2 | GPT-5.5 Supported OpenAI | 47.4% | source | — | — | |
| 3 | Claude Opus 4.7 Supported Anthropic | 26.5% | source | — | — | |
| 4 | Gemini 3.1 Pro Preview Supported Google | 26.1% | source | — | — | |
| 5 | Claude Sonnet 4.6 Supported Anthropic | 23.6% | source | — | — | |
| 6 | Kimi K2.6 Estimated Moonshot AI | 18.4% | source | — | — | |
| 7 | GLM-5.1 Estimated Z.ai (Zhipu) | 18.1% | source | — | — | |
| 8 | Claude Haiku 4.5 Estimated Anthropic | 13.7% | source | — | — | |
| 9 | MiniMax-M2.7 MiniMax | 13.3% | source | — | — |