Cybersecurity Benchmarks
AI security model leaderboards for vulnerability discovery, exploitation, and patching
Cybersecurity: Scores and Sources
Explore 5 benchmarks and their source-backed results below. Domain IQ uses the same shared scorer as the interactive charts and public API and remains outside Composite IQ. The cybersecurity scorer applies harness-aware adjustments; its domain IQ is not a raw benchmark score. Coverage shows how many domain benchmarks contribute to each estimate; incomplete coverage and different evaluation harnesses limit direct comparisons. Read the scoring methodology.
Highest domain scores among tracked public models
| Model | Domain IQ | Coverage |
|---|---|---|
| GPT 5.5 | 147 | 4/5 |
| GPT 5.6 Sol | 139 | 1/5 |
| GPT 5.4 | 138 | 4/5 |
| Glm 5.2 | 134 | 2/5 |
| Glm 5.3 | 134 | 1/5 |
| Minimax M3 | 134 | 1/5 |
| Deepseek V4 Pro | 133 | 2/5 |
| Grok 4.6 | 133 | 1/5 |
| Opus 4.6 | 131 | 3/5 |
| Deepseek V4 Flash | 127 | 2/5 |
| Gemini 3.5 Flash | 127 | 1/5 |
| Gemini 3.1 Pro | 121 | 4/5 |
| Opus 4.8 | 120 | 1/5 |
| Grok 4.3 | 119 | 1/5 |
| Kimi K2.6 | 119 | 2/5 |
CyberGym — 27 models
Success rate on CyberGym Level 1 vulnerability reproduction tasks. Higher scores are better. Source: CyberGym leaderboard.
| Model or system | Evaluation harness | Source-backed score |
|---|---|---|
| Opus 4.6 | Crystalline | 89.6% |
| Multi Model Mdash | MDASH | 88.45% |
| Fugu Cyber V1.0 | Sakana Fugu (provider-reported) | 86.9% |
| GPT 5.5 Cyber | OpenAI Agent | 85.6% |
| Glm 5.3 | Claude Code | 84.5% |
| Deepseek V4 Pro | DeepSeek Agent | 83.3% |
| Mythos 5 | Anthropic Agent | 83.1% |
| GPT 5.5 | OpenAI Agent | 81.8% |
| Grok 4.6 | Grok Build | 79.7% |
| GPT 5.4 | OpenAI Agent | 79% |
| Glm 5.2 | Claude Code | 77.2% |
| Deepseek V4 Flash | DeepSeek Agent | 76.7% |
| Glm 5.1 | Claude Code | 68.7% |
| Opus 4.6 | Anthropic Agent | 66.6% |
| Sonnet 4.6 | Anthropic Agent | 65.2% |
| Muse Spark 1.1 | Meta Agent | 59% |
| Deepseek V4 Pro | Claude Code | 57.7% |
| Opus 4.5 | Anthropic Agent | 50.63% |
| Muse Spark | Meta Agent | 43.5% |
| Glm 5 | Claude Code | 43.2% |
| Kimi K2.5 | Kimi Agent | 41.3% |
| GPT 5 | OpenHands | 39.4% |
| Sonnet 4.5 | Anthropic Agent | 28.9% |
| Opus 4.1 | Anthropic Agent | 25% |
| Sonnet 4 | Anthropic Agent | 22.6% |
| Sonnet 3.7 | Anthropic Agent | 14.5% |
| GPT 4.1 | OpenHands | 9.356% |
| Gemini 2.5 Flash | OpenHands | 4.844% |
| Deepseek V3 | OpenHands | 3.583% |
CyberBench — 13 models
Overall success across VALS CyberBench PoC and Patch tracks for OSS-Fuzz crash reproduction and repair. Higher scores are better. Source: VALS CyberBench.
| Model or system | Evaluation harness | Source-backed score |
|---|---|---|
| GPT 5.5 | mini-swe-agent | 80.508% |
| Glm 5.2 | mini-swe-agent | 77.359% |
| Minimax M3 | mini-swe-agent | 76.271% |
| Kimi K2.6 | mini-swe-agent | 74.351% |
| GPT 5.4 | mini-swe-agent | 72.881% |
| Gemini 3.5 Flash | mini-swe-agent | 69.331% |
| Deepseek V4 Pro | mini-swe-agent | 68.644% |
| Deepseek V4 Flash | mini-swe-agent | 66.716% |
| Opus 4.7 | mini-swe-agent | 52.141% |
| Opus 4.8 | mini-swe-agent | 50.847% |
| Grok 4.3 | mini-swe-agent | 47.457% |
| Qwen3.7 Plus | mini-swe-agent | 37.5% |
| Gemini 3.1 Pro | mini-swe-agent | 36.441% |
ExploitBench — 9 models
Cap coverage on the ExploitBench capability ladder for V8 exploitation. Higher scores are better. Source: ExploitBench leaderboard.
| Model or system | Evaluation harness | Source-backed score |
|---|---|---|
| Mythos 5 | AutoNudge | 78% |
| GPT 5.5 | Codex AutoNudge | 72% |
| Mythos 5 | Baseline | 72% |
| GPT 5.5 | AutoNudge | 47% |
| GPT 5.5 | AutoNudge | 42% |
| GPT 5.5 | Codex | 33% |
| Opus 4.7 | Baseline | 28% |
| Gemini 3.1 Pro | Baseline | 26% |
| Sonnet 4.6 | Baseline | 26% |
| Opus 4.7 | Baseline | 25% |
| Sonnet 4.6 | Baseline | 25% |
| Kimi K2.6 | AutoNudge | 18% |
| Glm 5.1 | AutoNudge | 18% |
| Kimi K2.6 | Baseline | 16% |
| Glm 5.1 | Baseline | 16% |
| Gemini 3.1 Pro | AutoNudge | 16% |
| Haiku 4.5 | AutoNudge | 14% |
| Haiku 4.5 | Baseline | 14% |
| Minimax M2.7 | Baseline | 13% |
| Minimax M2.7 | AutoNudge | 13% |
ExploitGym — 9 models
Intended-vulnerability exploits completed across userspace, V8, and Linux kernel tasks. Higher scores are better. Source: ExploitGym leaderboard.
| Model or system | Evaluation harness | Source-backed score |
|---|---|---|
| GPT 5.6 Sol | Codex CLI | 293 |
| Mythos 5 | Claude Code | 157 |
| GPT 5.5 | Codex CLI | 131 |
| GPT 5.4 | Codex CLI | 61 |
| Opus 4.6 | Claude Code | 16 |
| Gemini 3.1 Pro | Gemini CLI | 12 |
| Opus 4.7 | Claude Code | 12 |
| Muse Spark 1.1 | Meta Agent | 7 |
| Glm 5.1 | Claude Code | 4 |
CyberGym-E2E — 7 models
S3 end-to-end discover-and-patch success rate: PoC crashes, patch fixes it, and tests pass. Higher scores are better. Source: CyberGym-E2E leaderboard.
| Model or system | Evaluation harness | Source-backed score |
|---|---|---|
| GPT 5.4 | Codex | 65.9% |
| Opus 4.6 | Claude Code | 62.6% |
| Gemini 3.1 Pro | Gemini CLI | 43.8% |
| Opus 4.6 | Claude Code | 37.9% |
| Gemini 3 Pro | Gemini CLI | 22.6% |
| GPT 5.2 | Codex | 20.7% |
| Opus 4.5 | Claude Code | 19.2% |
| Sonnet 4.5 | Claude Code | 10.6% |
| Sonnet 4.5 | OpenHands | 5.4% |