Google Launches Gemini 4 Carbon, Scoring 88% on Security Vulnerability Benchmarks

Google has released Gemini 4 Carbon, a frontier AI model purpose-built for cybersecurity work, scoring 88.2% on SEC-bench Pro — the primary industry benchmark for software vulnerability identification. For security teams, this is the first externally accessible AI model specifically tuned for the sustained, technical work of finding and fixing code flaws rather than general-purpose question answering.

Google's earlier Gemini 4 Argon, released in October, targeted a narrow audience of defensive security practitioners. Carbon extends that lineage with significantly higher benchmark scores and a broader (though still restricted) access program. The distinction matters: Argon was an experiment; Carbon is a deployment.

Benchmark Performance

According to Google via The Neuron, Gemini 4 Carbon posted:

• 88.2% on SEC-bench Pro — identifying real software vulnerabilities in production-representative code
• 71% on CWE-bench — producing working exploit or fix code for identified Common Weakness Enumeration entries

For context: earlier general-purpose frontier models (GPT-4o, Claude 3.5 Sonnet) have scored in the 40–60% range on SEC-bench Pro. A score of 88.2% represents a substantial jump in reliable vulnerability detection, not an incremental one.

CWE-bench measures something harder: not just spotting a problem but generating code that either confirms exploitability or patches the flaw. A 71% success rate on working fixes is notable because it suggests Carbon can participate meaningfully in remediation workflows, not just triage.

Access: Restricted to Vetted Defenders

Carbon is not publicly available. Google is distributing it exclusively through its Fairwind program, which requires vetting of applicants as cyber defenders — security researchers, incident response teams, and authorized red teams. Broader availability is contingent on voluntary testing with U.S. government agencies, a process Google did not put a timeline on.

This restricted rollout reflects an explicit dual-use concern. A model that scores 88.2% on finding vulnerabilities is equally useful for offense as defense. Google's decision to gate access behind identity verification follows Anthropic's approach with Claude for cybersecurity use cases, but goes further by requiring active government review before general release.

What This Means for Security Teams

For enterprise security: A model that can reliably identify vulnerabilities at this accuracy rate changes the economics of code review. Manual security review of a large codebase is expensive and slow; AI-assisted triage at 88% accuracy could let small security teams cover substantially more surface area.

For the vulnerability research community: Carbon's CWE-bench performance means it can assist with proof-of-concept development — work that has historically required deep human expertise. Responsible disclosure timelines may compress if researchers can iterate faster on exploit validation.

For software vendors: The same capability that helps defenders find your bugs faster also means that anyone who gets unauthorized access to a Carbon-equivalent model could audit your code at scale. The window between vulnerability discovery and patch deployment is already a race; models like Carbon accelerate both sides.

The Argon–Carbon Distinction

Google releasing two named cybersecurity models in the same month — Argon and Carbon — signals a deliberate product line rather than a one-off experiment. Argon appears to have been a more conservative, narrower model for blue-team (defensive) use. Carbon extends into vulnerability research and fix generation, which requires broader access to offensive techniques. The naming convention (noble gases) suggests additional models in the series are likely.

What to Watch

Whether the U.S. government testing process produces a public report — that outcome would set a precedent for how AI cybersecurity models get certified before general release. Also watch for independent red-teaming of Carbon by the security research community once Fairwind access expands, which will stress-test whether Google's vetting is sufficient to prevent misuse.