Gemini 4 Carbon for cybersecurity

Security is the headline of Google's Gemini 4 Carbon announcement. Google calls its cybersecurity skills state of the art, and that is the reason defenders get it before anyone else. Here is what it can do, how it scored, and how security teams can get access.

How Carbon scores on security tests

Google tested three stages of a security job: finding a bug, proving it can be exploited, and fixing it. Carbon leads on all three.

Google ran ExploitGym and SEC-bench Pro itself, with one attempt per task and its own safeguards switched off, to measure the model's raw ability. Competitors' numbers come from their own published reports.

What partners report

• Chrome. Google's Chrome team says Carbon was more than 10 times better than another lab's large frontier model at writing high-quality patches. Google didn't name the other lab.
• Wiz. The cloud security company reports that Carbon leads its benchmarks for web penetration testing, part of its Scan for Good work.
• Secure code by default. On Google's internal test of writing secure code, Carbon's success rate is more than double Argon's.
• Long investigations. Google says Carbon can work for days on hard engineering jobs, like tracing a bug to its root cause across a large codebase and patching it.

Hard to trick, hard to misuse

A security model has to resist attacks on itself. Google reports two results on that front:

Prompt injection is when an attacker hides instructions inside a web page or document to hijack an AI agent. On the Gray Swan test, nothing got through Carbon; Argon was at 0.7%, Claude Opus 5.5 at 1.0% and GPT-6 Astra at 8.5%.

Carbon also runs inside the same layered safety system as Argon: safety training, real-time misuse detection, prompt-injection defenses and monitoring of the model's reasoning. Google says it trained Carbon to handle complex or high-pressure tasks appropriately, and it is taking part in the US government's voluntary testing before wider release.

The Fairwind Program: how defenders get access

Fairwind is Google DeepMind's program that gives vetted defenders early access to its most capable models, so they can fix weaknesses before attackers find them.

• Who it's for: governments and national cyber agencies, critical infrastructure operators (healthcare, telecoms, energy, finance), major technology platforms, and academic labs focused on defensive testing.
• How to join: fill in the interest form on the Fairwind page. Google reviews applicants, including their security history, and replies to eligible ones.
• The rules: use is limited to internal security, incident response and penetration testing teams, with strong sign-in protection. Permitted work includes authorized threat simulation, reverse engineering and malware analysis for defense or research. Creating malware is not allowed, and access can't be shared or resold.
• Not eligible? Google points other teams to CodeMender, its code security agent, which works with publicly available models.

Apply at deepmind.google/fairwind-program. We have no connection to the program and can't help with applications.

What security teams can do today

If you're not in Fairwind, Gemini 4 Argon is available now in the Gemini API at the same price as Carbon, and it was already first or tied for first on CWE-bench. Build your review and patching workflows on Argon, keep the model name in one setting, and switch when Carbon opens. Our API guide has the code.

Cybersecurity questions

Sources: Google's Carbon announcement, evaluation methodology, Fairwind Program.