Google Says Gemini 4 Carbon Leads Rivals on Vulnerability Benchmarks
Summary
Google says its unreleased Gemini 4 Carbon topped rival models on vulnerability-discovery, exploitation and patching benchmarks. The company is opening staged access to selected cyber defenders and reports additional coding and safety gains.
Key points
- Google says Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear's CWE-bench, outperforming rival models in its comparisons.
- Google says the benchmark results were obtained without guardrails.
- Access is opening to a select group of cyber defenders, with staged expansion to approved partners through the Fairwind Program; the model is not broadly available yet.
- Google says Carbon was more than twice as successful as Gemini 4 Argon at writing secure code; its Chrome team reported Carbon produced high-quality patches at more than 10 times the rate of a rival model.
- Google reports no successful attacks against Carbon in Gray Swan's indirect prompt injection benchmark, compared with a 0.7% success rate for Argon.
- Google says internal teams saw nearly 80% fewer high-severity misalignment alerts when using Carbon in real coding environments.
Article Details
- Event Type
- Google announced Gemini 4 Carbon, a cybersecurity-focused AI model, and began a staged limited-access rollout.
- Impact
- Google says Carbon outperformed rival models on vulnerability discovery, exploitation, and patching benchmarks, and reported improved safety results compared with Gemini 4 Argon. The benchmark results and safety claims are attributed to Google and the named benchmark evaluations. Access is initially limited to selected cyber defenders and will expand to approved partners.
Vendors
Anthropicbenchmark and 71.0% on Collinear’s CWE-bench patching benchmark, topping OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in the company’s comparison.GoogleGoogle on Wednesday announced Gemini 4 Carbon, a frontier AI model that, according to the company, beats rival models on benchmarks for detecting, exploiting and patching vulnerabilities.OpenAIon the ExploitGym exploitation benchmark and 71.0% on Collinear’s CWE-bench patching benchmark, topping OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in the company’s comparison.
Products
Claude Fable 5.1CWE-bench patching benchmark, topping OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in the company’s comparison.Claude Opus 5.5and 71.0% on Collinear’s CWE-bench patching benchmark, topping OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in the company’s comparison.Gemini 4 ArgonCarbon follows Gemini 4 Argon, which Google launched two weeks ago.Gemini 4 CarbonGoogle on Wednesday announced Gemini 4 Carbon, a frontier AI model that, according to the company, beats rival models on benchmarks for detecting, exploiting and patching vulnerabilities.GPT-6.1 Astraexploitation benchmark and 71.0% on Collinear’s CWE-bench patching benchmark, topping OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in the company’s comparison.