ExploitBench V8 leaderboard
As of November 22, 2026, among the results listed here, OpenAI GPT-6 Astra has the highest reported ExploitBench V8 score, 100.0% (Mean capability credit (five attempts per vulnerability)), published by OpenAI. 5 models have a ExploitBench V8 result in our records.
Scores as the model makers published them, not re-run by us; each row links to its announcement.
Best reported score
100.0%
Reported by OpenAI, November 17, 2026 · Mean capability credit (five attempts per vulnerability)
Tied for first
100.0%
Reported by OpenAI, November 17, 2026 · Mean capability credit (five attempts per vulnerability)
5 models ranked · 1 maker
| Rank | Model | Score | Setting | Reported by | Date |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 100.0% | Mean capability credit (five attempts per vulnerability) | OpenAI | November 17, 2026 (model release) |
| 1 | GPT-6.2 Astra | 100.0% | Mean capability credit (five attempts per vulnerability) | OpenAI | November 17, 2026 (model release) |
| 1 | GPT-6.2 Sol | 100.0% | Mean capability credit (five attempts per vulnerability) | OpenAI | November 17, 2026 (model release) |
| 4 | GPT-6.1 Sol | 99.7% | Mean capability credit (five attempts per vulnerability) | OpenAI | November 17, 2026 (model release) |
| 5 | GPT-6.1 Terra | 97.1% | Mean capability credit (five attempts per vulnerability) | OpenAI | November 17, 2026 (model release) |
How much to trust these numbers
- Source
- Reported by the model makers themselves. Not re-run by LLMBenchmark or anyone independent.
- Settings
- All ranked results use the same reported setting.
- Spread
- 100.0% to 97.1% across 5 models.
- Freshness
- Published November 17, 2026.
What ExploitBench V8 measures
It asks models to exploit patched V8 tasks; the score indicates how many of 16 capability flags they achieve, from coverage and crashes to arbitrary code execution. Source: ExploitBench