ExploitBench
ReportedNo description yet: we only describe an eval set when we can do it correctly.
- Kind
- Published by a third party's leaderboard, collected by Epoch AI. The leaderboard is named and linked.
- Score shown
- Epoch's column "Mean capability", as a percentage, higher is better; random guessing scores about 0.0%. The source gives no error bars.
- Results
- 20 in all: 6 for open-weight models, 14 for closed ones. 6 linked to a model page here.
Best open-weight result: moonshotai/Kimi-K2.6, 18.4%, number 12 of 20 results.
| # | Model | Weights | Setting | Score | Date | Source |
|---|---|---|---|---|---|---|
| 1 | claude-mythos-preview-earlyAnthropic | closed | default | 73.8% | not dated | via Epoch AI |
| 2 | claude-mythos-preview-earlyAnthropic | closed | default | 70.6% | not dated | via Epoch AI |
| 3 | gpt-5.5_unknownOpenAI | closed | effort unknown | 47.4% | not dated | via Epoch AI |
| 4 | gpt-5.5_unknownOpenAI | closed | effort unknown | 41.8% | not dated | via Epoch AI |
| 5 | gpt-5.5_unknownOpenAI | closed | effort unknown; Harness: Codex | 41.6% | not dated | via Epoch AI |
| 6 | gpt-5.5_unknownOpenAI | closed | effort unknown; Harness: Codex | 32.6% | not dated | via Epoch AI |
| 7 | claude-opus-4-7Anthropic | closed | default | 26.5% | not dated | via Epoch AI |
| 8 | gemini-3.1-pro-previewGoogle DeepMind | closed | default | 26.1% | not dated | via Epoch AI |
| 9 | claude-opus-4-7Anthropic | closed | default | 23.9% | not dated | via Epoch AI |
| 10 | claude-sonnet-4-6Anthropic | closed | default | 23.6% | not dated | via Epoch AI |
| 11 | claude-sonnet-4-6Anthropic | closed | default | 22.7% | not dated | via Epoch AI |
| 12 | moonshotai/Kimi-K2.6Epoch: kimi-k2.6, Moonshot | open | default | 18.4% | not dated | via Epoch AI |
| 13 | zai-org/GLM-5.1Epoch: glm-5.1, Z.ai (Zhipu AI) | open | default | 18.1% | not dated | via Epoch AI |
| 14 | moonshotai/Kimi-K2.6Epoch: kimi-k2.6, Moonshot | open | default | 16.3% | not dated | via Epoch AI |
| 15 | zai-org/GLM-5.1Epoch: glm-5.1, Z.ai (Zhipu AI) | open | default | 16.0% | not dated | via Epoch AI |
| 16 | gemini-3.1-pro-previewGoogle DeepMind | closed | default | 15.7% | not dated | via Epoch AI |
| 17 | claude-haiku-4-5-20251001_unknownAnthropic | closed | effort unknown | 13.7% | not dated | via Epoch AI |
| 18 | claude-haiku-4-5-20251001_unknownAnthropic | closed | effort unknown | 13.4% | not dated | via Epoch AI |
| 19 | MiniMaxAI/MiniMax-M2.7Epoch: MiniMax-M2.7, MiniMax | open | default | 13.3% | not dated | via Epoch AI |
| 20 | MiniMaxAI/MiniMax-M2.7Epoch: MiniMax-M2.7, MiniMax | open | default | 13.3% | not dated | via Epoch AI |
Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method