Skip to content

Benchmark

ExploitGym on-target exploits with mitigations (V8)

Higher scores are better.

Models measured
8
Scores recorded
9
Sources
1
IQ weight
0.65

Score distribution

6.1
Mean
0
Median
10.9
Std dev
0
10th percentile
19
90th percentile
0
Min
032
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.65 for chat models, measured over 8 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1GPT-5.6 Solagent: Codex CLI, model note: reasoning max, eval: 6h timeout, version: v1, focus: model, reported by: OpenAI & ExploitGym Team, date: 2026-07-13· ExploitGym leaderboard data docs32
2GPT-5.6 Solagent: Codex CLI, model note: reasoning max, eval: 2h timeout cutoff, version: v1, focus: model, reported by: OpenAI & ExploitGym Team, date: 2026-07-13· ExploitGym leaderboard data docs19
3GPT 5.5agent: Codex CLI, model note: OpenAI's default safety filters block all GPT-5.5 exploit attempts under default prompting., eval: 2h timeout, version: v1, focus: model, reported by: ExploitGym Team, date: 2026-06-16· ExploitGym leaderboard data docs4
4Claude Opus 4.6agent: Claude Code, model note: Results obtained in collaboration with Anthropic., eval: 2h timeout, version: v1, focus: model, reported by: ExploitGym Team, date: 2026-06-16· ExploitGym leaderboard data docs0
5Claude Opus 4.7agent: Claude Code, eval: 2h timeout, version: v1, focus: model, reported by: ExploitGym Team, date: 2026-06-16· ExploitGym leaderboard data docs0
6Gemini 3.1 Pro Previewagent: Gemini CLI, eval: 2h timeout, version: v0, focus: model, reported by: ExploitGym Team, date: 2026-05-13· ExploitGym leaderboard data docs0
7GLM 5.1agent: Claude Code, eval: 2h timeout, version: v1, focus: model, reported by: ExploitGym Team, date: 2026-06-16· ExploitGym leaderboard data docs0
8GPT 5.4agent: Codex CLI, eval: 2h timeout, version: v1, focus: model, reported by: ExploitGym Team, date: 2026-06-16· ExploitGym leaderboard data docs0
9Muse Spark 1.1agent: Meta Agent (off-the-shelf agent harness), model note: helpful-only versoin, eval: 4h timeout, version: v1, focus: model, reported by: Meta AI, date: 2026-07-09· ExploitGym leaderboard data docs0

Data credits: ExploitGym leaderboard (Wang et al., UC Berkeley), arXiv:2605.11086. No licence is stated for the result data; the repository's Apache-2.0 licence covers its code only, and task data keeps its upstream licences. Literal exploit counts; hidden rows omitted.