Google 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.
The model is not broadly available yet. This week, Google is opening access to a select group of cyber defenders, then expanding it in stages to a wider pool of approved partners via its Fairwind Program.
Fairwind offers governments, critical infrastructure operators and major technology platforms early access to advanced models for cyber defense. The company says more than 650 organizations currently participate.
Carbon follows Gemini 4 Argon, which Google launched two weeks ago. Some Fairwind partners already have exclusive access to Argon’s cybersecurity defense capabilities.
Google said Carbon scored 88.2% on SEC-bench Pro, which tests a model’s ability to uncover flaws in real-world software. It also achieved 54.8% on 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.
These results were obtained with no guardrails in place, the company noted.
Google frames exploitation as a defensive step that helps teams decide what to fix. “To prioritize which vulnerabilities to remediate first or remediate at all, cyber defenders increasingly rely on models to check which vulnerabilities are exploitable,” the company said.
Based on Google’s internal benchmarks, Carbon is more than twice as successful as Argon at writing secure code. When it came to generating high-quality patches, Google’s Chrome team found Carbon outperformed a large frontier model from a rival lab by more than 10 times.
Wiz, the Google-owned cloud security firm, has observed Carbon taking the top spot on web penetration testing benchmarks.
On the safety front, no attack succeeded against Carbon in Gray Swan’s indirect prompt injection benchmark, compared with a 0.7% success rate for Argon. On Terminal-Bench 4.0, none of Carbon’s cheating attempts succeeded, and those attempts fell to 0.4%.
Compared with Argon, the new model made about half as many unsupported claims. Google’s internal teams, using Carbon in real coding environments, saw a nearly 80% decrease in high-severity misalignment alerts.
Google says it is involved in the US government’s voluntary effort to test models before they are released. “We’re continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible,” the company said.
Carbon will launch at Argon’s introductory rates of $2 per million input tokens and $10 per million output tokens. When the introductory period ends, the price goes up to $4 and $20, respectively.
Related: Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps
Related: Stolen AI Reasoning Could Help China Build Cheaper Rivals
Related: Mythos 5 Ignored Real-World Harm in July Attack
