ExploitBench v8-bench (ExploitBench)
We show this table for reference; we do not rank on it.
A cybersecurity benchmark for evaluating LLM agents on full-control V8 exploit synthesis using 16 measured exploit capability flags.
Capability coverage on ExploitBench — Current v8-bench aggregate table
We mirror the published capability coverage view for ExploitBench. Claude Mythos Preview leads the public snapshot at 78%, followed by GPT 5.5 (Codex) (72%) and Claude Mythos Preview (72%). We do not use these results to rank models overall.
Claude Mythos Preview
Anthropic
AutoNudge, 5 seeds.
GPT 5.5 (Codex)
OpenAI
Codex
AutoNudge, 3,000-turn budget.
Claude Mythos Preview
Anthropic
5 seeds.
20 modelsAgenticCurrentDisplay onlyUpdated Current v8-bench aggregate table
Capability coverage table (20 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Mythos PreviewAnthropicAutoNudge, 5 seeds. | 78% | Not reported | Closed |
| 2 | GPT 5.5 (Codex)OpenAICodexAutoNudge, 3,000-turn budget. | 72% | Not reported | Closed |
| 3 | Claude Mythos PreviewAnthropic5 seeds. | 72% | Not reported | Closed |
| 4 | GPT 5.5OpenAI | 47% | Not reported | Closed |
| 5 | GPT 5.5OpenAI | 42% | Not reported | Closed |
| 6 | GPT 5.5 (Codex)OpenAICodex | 33% | Not reported | Closed |
| 7 | Claude Opus 4.7AnthropicAutoNudge, 5 seeds. | 28% | Not reported | Closed |
| 8 | Gemini 3.1 Pro PreviewGoogle | 26% | Not reported | Unknown |
| 9 | Claude Sonnet 4.6AnthropicAutoNudge, 5 seeds. | 26% | Not reported | Closed |
| 10 | Claude Opus 4.7Anthropic5 seeds. | 25% | Not reported | Closed |
| 11 | Claude Sonnet 4.6Anthropic5 seeds. | 25% | Not reported | Closed |
| 12 | Kimi K2.6Moonshot AI | 18% | Not reported | Open |
| 13 | Glm 5.1Z.AI | 18% | Not reported | Open |
| 14 | Kimi K2.6Moonshot AI | 16% | Not reported | Open |
| 15 | Glm 5.1Z.AI | 16% | Not reported | Open |
| 16 | Gemini 3.1 Pro PreviewGoogle | 16% | Not reported | Unknown |
| 17 | Claude Haiku 4.5AnthropicAutoNudge, 5 seeds. | 14% | Not reported | Closed |
| 18 | Claude Haiku 4.5Anthropic5 seeds. | 14% | Not reported | Closed |
| 19 | MiniMax M2.7MiniMax | 13% | Not reported | Open |
| 20 | MiniMax M2.7MiniMax | 13% | Not reported | Open |
How ExploitBench is shown here
BenchLM mirrors the public ExploitBench v8-bench leaderboard. ExploitBench evaluates LLM cybersecurity agents on V8 exploit synthesis and reports coverage over 16 exploit-capability flags.
ExploitBench is display only on BenchLM because the public rows combine model, environment, and AutoNudge or harness settings. BenchLM keeps the official rows as security-evaluation context rather than weighted model-only evidence.
Snapshot
The published ExploitBench snapshot places Claude Mythos Preview first at 78%. The third row is 6 points behind. The broader top-10 range is 53 points, so the table still separates the published systems.
20 models have been evaluated on ExploitBench. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. ExploitBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About ExploitBench
Year
2026
Tasks
V8 exploit synthesis runs
Format
Capability coverage percentage over 16 flags
Difficulty
Browser exploitation and cybersecurity
ExploitBench measures whether LLM agents can turn patched V8 bugs into progressively stronger exploit capabilities, from reaching vulnerable code to full control. BenchLM mirrors the official public leaderboard as display-only security-evaluation context.
Freshness and provenance
Version
ExploitBench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does ExploitBench measure?
A cybersecurity benchmark for evaluating LLM agents on full-control V8 exploit synthesis using 16 measured exploit capability flags.
Which model leads the published ExploitBench snapshot?
Claude Mythos Preview currently leads the published ExploitBench snapshot with 78% capability coverage. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on ExploitBench?
The Current v8-bench aggregate table snapshot contains 20 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.