ExploitBench
Integrity rank #17 of 65 · 8 models scored · top score 47.4 · GPT-5.5
unknown
Strongest on contamination resistance, weakest on discrimination. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | GPT-5.5 | 47.4 | unverifiedT2 | 2026-04-23 |
| 2 | Claude Opus 4.7 | 26.5 | unverifiedT2 | 2026-04-16 |
| 3 | Gemini 3.1 Pro | 26.1 | unverifiedT2 | 2026-02-19 |
| 4 | Claude Sonnet 4.6 | 23.6 | unverifiedT2 | 2026-02-17 |
| 5 | Kimi K2.6 | 18.4 | unverifiedT2 | 2026-04-20 |
| 6 | GLM-5.1 | 18.1 | unverifiedT2 | 2026-04-07 |
| 7 | Claude Haiku 4.5 | 13.7 | unverifiedT2 | 2025-10-15 |
| 8 | MiniMax-M2.7 | 13.3 | unverifiedT2 | 2026-03-18 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 8 models, and the gap between its best and worst performers spans 34.1 points. A spread that wide supports genuine separation between models, so rank differences are more likely to reflect real capability gaps than noise.
3 cited facts
The benchmark has an integrity score of 94 (grade A), ranking 15th among 62 benchmarks, reflecting its health as a discriminator of frontier models under a disclosed harness. The weakest component is discrimination, which means the benchmark's limited ability to separate frontier models makes fine-grained ranking unreliable.
5 cited facts
This benchmark is not saturated, with a top score of 47.4 points and 52.6 points of headroom remaining to the ceiling. Only one model sits near the top, so there is no broad cluster band within a couple of points. That means there is still genuine room to separate the field at the top, rather than the top ranking being close to noise.
5 cited facts
Under the consistent harness reported for this benchmark, scores are directly comparable as task performance under a disclosed evaluation setup, and the harness notes identify the model, model version, performance, optimized flag, score source, release date, and model date used in that reporting. The test set privacy is unknown, so it is not possible to say whether the test set is public or held out and thus whether high scores should receive the additional contamination scrutiny that a public set would warrant. Contamination history is also unknown, and the benchmark should be read as measuring task performance under a disclosed harness rather than deployed capability; if comparability were low, scores from different harnesses should not be compared head-to-head.
5 cited facts
What this benchmark actually measures is not documented in the available record, so its target quantity and the behavior it is meant to score cannot be stated here. The construction of the test — how tasks were assembled, filtered, or validated — is likewise not documented, leaving the pipeline behind the benchmark undescribed. Its key assumptions and its acknowledged limitations are both unrecorded, so no principled statement can be made about the conditions under which results should be trusted. Because the limitations are undocumented, the sharpest caveat a reader should hold onto cannot be recovered from these fields, and no human baseline is recorded either, so there is no documented reference point for judging model performance.
6 cited facts
ExploitBench: ExploitBench as reported in Epoch AI's Capabilities Index CSV.
GPT-5.5 leads ExploitBench at 47.4. The full leaderboard above lists every recorded measurement, not just the headline number.
8 models have recorded scores on ExploitBench, spanning a score spread of 34.1.
tensor.news grades ExploitBench A for integrity (score 94/100), ranking #17 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — ExploitBench still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on ExploitBench are reasonably apples-to-apples.
Every ExploitBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.