Best AI models for cybersecurity
Google reports Gemini 4 Carbon at 71% on CWE-bench v1, the highest of the 12 models with a result here, with 0% attack success on Gray Swan IPI at 15 attempts. These results stay separate from earlier CWE-Bench results and CyberGym: the tasks, harnesses, and safeguards differ, so they do not support one combined cyber ranking.
Use this page to compare like-for-like results, inspect the source behind each number, and separate model capability from trusted-access policy.
Vulnerability defense
Gemini 4 Carbon
Google reports 71% patching success on CWE-bench v1, the highest of the 12 models with a result here; Gemini 4 Argon follows at 68%.
Broadest coverage
GPT-6 Astra
Its rows in the comparison below span CWE-bench v1, Gray Swan IPI, ExploitGym, ExploitBench, SEC-Bench Pro, FrontierCyber, ExploitGym Hardened, ExploitBench Internal Port, Web Pentesting and ExploitGym Honeypot.
ExploitBench
GPT-6 Cyber: 100%
Tied for the highest of 16 reported ExploitBench results; GPT-6 Astra follows at 100%. Labs run it with their own harnesses and safeguard settings.
Trusted access
Blue is access; Red is a model
Daybreak Blue changes safeguards around GPT-5.6 Sol. Daybreak Red exposes the separate GPT-5.6 Cyber and GPT-6 Cyber models.
Cyber benchmark comparison
Dashes mean no exact public result is attached to that model and protocol.
| Model | CyberGym | CWE-Bench | CWE-bench v1 | Gray Swan IPI (lower is better) | ExploitGym | ExploitBench | SEC-Bench Pro | The Last Ones | FrontierCyber | ExploitGym Hardened | ExploitBench Internal Port | RemoteExploitBench | Web Pentesting | ExploitGym Honeypot (lower is better) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-6 Astra OpenAI | — | — | 68% | 8.5% | 42.4% | 100% | 85.4% | — | 38.1% | 29.7% | 39% | — | 32% | 0% |
| GPT-5.6 Sol OpenAI | 84.5% | — | — | 27% | 33.7% | 73.5% | 71.2% | 70% | 9.6% | — | 5.5% | — | — | 49.5% |
| GPT-6 Cyber OpenAI | — | — | — | — | 48.5% | 100% | 87% | — | — | 36.9% | 42% | 94.7% | 40.3% | 0.07% |
| GPT-6 Sol OpenAI | — | — | 52% | 10.1% | 22.1% | — | 66.3% | — | — | — | 5.5% | — | — | 0% |
| GPT-6.1 Sol OpenAI | — | — | — | — | 35.1% | 99.7% | 78.8% | — | — | — | 21.5% | — | — | 0% |
| Gemini 4 Carbon | — | — | 71% | 0% | 53.5% | — | 88.2% | — | — | — | — | — | — | — |
| GPT-6.1 Terra OpenAI | — | — | — | — | — | 97.1% | 83.33% | — | — | — | 24.5% | — | — | 0% |
| DeepSeek V4.1 Flash 1017 DeepSeek | 88.1% | — | 55% | — | 15.3% | — | 62.8% | — | — | — | — | — | — | — |
| Gemini 4 Argon | — | — | 68% | 0.7% | 39.5% | — | 83.4% | — | — | — | — | — | — | — |
| MiMo-V2.6-Pro Xiaomi | 94% | — | — | — | 17.8% | 47.9% | 66.3% | — | — | — | — | — | — | — |
| MiMo-V2.6-Flash Xiaomi | 95.1% | — | — | — | 6% | 25.3% | 47.5% | — | — | — | — | — | — | — |
| GLM-5.3 Z.AI | 84.5% | — | — | 31.5% | 15% | 54.4% | — | — | — | — | — | — | — | — |
| GPT-5.6 Luna OpenAI | 77.9% | — | — | — | 12.4% | 33.2% | — | — | — | — | — | — | — | 9.9% |
| GPT-5.5 OpenAI | 81.8% | — | — | — | 13.4% | — | 45.8% | — | — | 2.3% | — | — | — | — |
| Claude Mythos Preview Anthropic | 83.1% | — | — | — | 17.5% | 69% | — | — | — | 5% | — | — | — | — |
| GPT-6 Luna OpenAI | — | — | — | — | 11.6% | — | 34.2% | — | — | — | 0% | — | — | 0% |
| Claude Opus 5.5 Anthropic | — | — | 67% | 1% | 33.3% | — | — | — | — | — | — | — | — | — |
| Gemini 3.8 Flash Cyber | 86.2% | 47.2% | — | 6% | — | — | — | — | — | — | — | — | — | — |
| Kimi K3 Moonshot AI | — | — | — | 52.7% | — | 32% | — | 10% | — | — | — | — | — | — |
| GPT-5.6 Terra OpenAI | 81.8% | — | — | — | 23.2% | 52.9% | — | — | — | — | — | — | — | — |
| Claude Mythos 5 Anthropic | 83.8% | — | — | — | — | 78% | — | 60% | — | — | — | — | — | — |
| GPT-5.4 OpenAI | 79% | — | — | — | 6% | — | — | — | — | 0.3% | — | — | — | — |
| Claude Mythos 5.2 Anthropic | — | — | — | — | 38.1% | 95% | — | — | — | — | — | — | — | — |
| Kimi K3.1 Pro Moonshot AI | — | — | — | — | 20.7% | 74.7% | — | — | — | — | — | — | — | — |
| Muse Spark 1.3 Meta | — | — | 55% | 15.9% | — | — | — | — | — | — | — | — | — | — |
| Claude Fable 5.1 Anthropic | — | — | 58% | 1% | — | — | — | — | — | — | — | — | — | — |
| Hy4 preview Tencent | 78.4% | — | 53% | — | — | — | — | — | — | — | — | — | — | — |
| Grok 4.6 xAI | — | — | 57% | 51.8% | — | — | — | — | — | — | — | — | — | — |
| GPT-5.6 Cyber OpenAI | — | — | — | — | 32.1% | — | — | — | — | — | — | — | — | 5.51% |
| Muse Spark 1.1 Meta | 59% | — | — | — | 0.8% | — | — | — | — | — | — | — | — | — |
| SWE-2 Flash Cognition | — | — | — | — | — | 34.8% | — | — | — | — | — | — | — | — |
| MiniMax M3.1 Flash MiniMax | 82.2% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Ling 3.1 Flash InclusionAI | 87.9% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Grok 4.7 xAI | — | — | 68% | — | — | — | — | — | — | — | — | — | — | — |
| Step 5 Preview StepFun | 84.7% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Atria Dawn Preview Shanghai Artificial Intelligence Laboratory | 86.5% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Gemini 3.8 Flash | — | — | — | 5.5% | — | — | — | — | — | — | — | — | — | — |
| Claude Mythos 5.1 Anthropic | — | — | — | — | 23.4% | — | — | — | — | — | — | — | — | — |
| DeepSeek V4 Pro 0813 DeepSeek | 83.3% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| DeepSeek V4 Flash 0731 DeepSeek | 76.7% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Claude Opus 5 Anthropic | — | — | — | 4.6% | — | — | — | — | — | — | — | — | — | — |
| Gemini 3.5 Flash Cyber | 83.2% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Fugu Cyber Sakana AI | 86.9% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Inkling Thinking Machines Lab | — | — | 37% | — | — | — | — | — | — | — | — | — | — | — |
| Claude Opus 4.7 (Adaptive) Anthropic | 73.1% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Muse Spark Meta | 43.5% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| GLM-5.1 Z.AI | 68.7% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| GLM-5 Z.AI | 43.2% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Claude Sonnet 4.6 Anthropic | 65.2% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Claude Opus 4.6 Anthropic | 66.6% | — | — | — | — | — | — | — | — | — | — | — | — | — |
| Claude Opus 4.5 Anthropic | 50.6% | — | — | — | — | — | — | — | — | — | — | — | — | — |
These columns are not averaged. CWE-bench v1 stays separate from earlier CWE-Bench results. Gray Swan IPI measures attack success at 15 attempts; lower is better. ExploitGym Honeypot measures how often a model exploits an out-of-scope service; lower is better. RemoteExploitBench, Web Pentesting and the ExploitBench internal port are OpenAI-run evaluations. Provider-run, independent, private, long-horizon, and access-compliance results remain separate.
Cybersecurity benchmark directory
Open a benchmark profile for its protocol, source, model rows, and verification status.
CyberBench
Can autonomous agents craft PoC inputs that trigger OSS-Fuzz vulnerabilities—and stop crashing after the fix?
Accuracy score · display only
Evaluation results published by Vals AI.
How to read the results
Match protocols. Compare CyberGym with CyberGym and ExploitBench with ExploitBench. Do not average an exploit success rate with a cyber-range step count.
Check the harness. Agent scaffolds, time limits, token budgets, source-code access, and Best@k settings can move the result as much as the model.
Separate access from capability. Advanced Cyber Completion Rate measures whether approved requests are completed rather than refused. A high ACCR is not a high cyber skill score.
Treat private evaluations cautiously. Provider system cards add useful evidence, but independent and benchmark-native replications carry a different evidentiary weight.
Daybreak access map
- Standard
- GPT-5.6 Sol with standard safeguards.
- Daybreak Blue
- The same GPT-5.6 Sol model through the
gpt-daybreak-bluealias, with modified safeguards for approved users. - Daybreak Red
- The separate GPT-5.6 Cyber model through the
gpt-daybreak-redalias, and GPT-6 Cyber (gpt-6-cyber, a cyber-specialized GPT-6 Astra) for approved Red customers.
Questions about AI cybersecurity benchmarks
What is the best AI model for cybersecurity?
There is no defensible single winner across the public evidence. GPT-6 Astra has the broadest mix of cyber results in this snapshot (10 of the comparison benchmarks), while GPT-6 Cyber reports a share of the highest ExploitBench result (100, against 100 for GPT-6 Astra). Access, safeguards, harnesses, and task types differ, so compare models within the same benchmark protocol.
Are Daybreak Blue and Daybreak Red separate models?
Daybreak Blue is an access configuration for GPT-5.6 Sol with modified cyber safeguards. Daybreak Red provides trusted access to separate cyber-specialized models: GPT-5.6 Cyber and, since November 5, 2026, GPT-6 Cyber, which OpenAI is opening to a select group of Red customers first.
Why is there no combined cybersecurity score?
CyberGym, CWE-Bench, ExploitGym, ExploitBench, FrontierCyber, SCONE, and ACCR measure different things under different harnesses. Averaging them would hide protocol differences and could turn an access-refusal metric into a capability claim.