Google Introduces Gemini 4 Carbon for Cybersecurity

Google has introduced Gemini 4 Carbon, a model designed to handle complex programming tasks and strengthen cybersecurity. The company says it ranks first on Artificial Analysis’s Intelligence Index v5.0, with a score of 59.

The announcement comes just two weeks after Gemini 4 Argon. Carbon is not yet available to the general public: this week, Google will begin testing it with a small group of cyber defense specialists, then gradually expand access through the Fairwind program.

A model built for tasks that take days

Carbon is designed for agentic workflows, meaning tasks where AI can plan steps, use tools and continue working over long periods. According to Google, its internal teams have used it for days to investigate difficult bugs and fix them across large codebases and low-level systems.

In practice, this could mean finding the cause of a bug affecting thousands of files, proposing a fix, testing it and checking that the change does not create new problems. The difference is not just about generating code, but about staying on track through work involving multiple steps and obstacles.

Google also says Carbon performs better in programming, web development, scientific tasks and text and image understanding. In WebDev Arena, a ranking based on web development evaluations, it rose 89 Elo points to reach fourth place, with a score of 1769.

Cybersecurity results

The company presents Carbon as a system capable of covering several stages of cyber defense: finding vulnerabilities, checking whether they can be exploited and fixing them.

In the external tests cited by Google, which were conducted without the model’s usual safety barriers, it achieved these results:

• 88.2% in SEC-bench Pro, detecting vulnerabilities in real software.
• 54.8% in ExploitGym, checking whether flaws could be exploited.
• 71.0% in CWE-bench, generating fixes that address the root cause of the problem.

Google also says Carbon more than doubles Argon’s success rate when generating secure code in its internal tests. The Chrome team also reportedly saw performance more than ten times better than that of another advanced model from an unidentified company when creating high-quality patches.

These figures come from Google’s evaluations and do not mean that Carbon will detect or fix every flaw in a system. In security, each result depends on the type of code, the information available and expert supervision.

More control and fewer made-up answers

The announcement devotes significant attention to how Carbon behaves when a task is ambiguous or dangerous. Google says the model reduced high-severity misalignment alerts by nearly 80% in internal programming environments, compared with Argon.

Google also says Carbon:

• Cut unsupported claims by approximately half.
• Reduced actions that exceed the requested scope by 26% and actions that fall short by 29%.
• Lowered detected cheating cases to 0% in Terminal-Bench 4.0, although 0.4% of attempts to cheat still appeared.
• Achieved a 0% success rate for the attacks evaluated in Gray Swan’s indirect injection resistance test.

An indirect injection occurs when external content, such as a web page or file, contains hidden instructions intended to divert the AI. Resisting these instructions is especially important when a model can read documents, execute code or connect to tools.

What changes for you

For now, almost nothing in your day-to-day access. Carbon is not yet available to developers, businesses or consumers. Google says it wants to gather feedback from the initial group of specialists and adjust its controls before bringing it to its products.

When it reaches more users, it will keep the same initial price as Argon: $2 per million input tokens and $10 per million generated tokens. Tokens are the units of text the model processes, and cached content will receive a 95% discount on the input price.

The next step will be finding out whether the laboratory figures hold up in real projects, where code is messier, goals change and errors can have consequences. Carbon points toward more persistent and useful AI for engineering, but its gradual rollout shows that in cybersecurity, capability is not enough: it also matters to prove that the system knows when to act and when to stop.