New LLM: Gemini 4 Carbon Is an Exploit Expert

What happened

Google has announced Gemini 4 Carbon, its newest large language model (LLM). An LLM is an AI system that reads and writes text. Carbon replaces Argon, which Google showed only four weeks ago. For now, Carbon is not open to everyone. Google says it will first share the model with US intelligence agencies and selected users. Over time, chosen cyber defenders and other trusted partners will get access through a program called Fairwind.

Key numbers

• On the Exploitgym test, Carbon scores almost 55 percent. Argon scored 42 percent.
• On the SEC-bench Pro test, Carbon scores just over 88 percent, ahead of GPT-6 Astra.
• On the CWE-bench test for fixing security holes, Carbon scores 71 percent.
• Price: 4 US dollars per million input tokens and 20 US dollars per million output tokens. A 50 percent introductory discount applies at first.
• Google has not said when Carbon will be public. It says this will happen soon. Argon took about two weeks to become public.

Why the access is limited

The article says Google is following a request from the White House. The company says it will first limit new AI models to US intelligence services and users they approve. Google also says no prompt injection attack has worked on Carbon so far. Prompt injection is a trick where hidden instructions in a text try to make the AI do something its user did not ask for.

Why it matters

Carbon's biggest gain is in finding and writing attacks on software, called exploits. This is a skill that can help security experts protect systems. It can also be misused by attackers. That is why Google is keeping access narrow. The story shows how AI companies are balancing fast progress with safety and government rules.

What to watch

• When Carbon becomes available to the public.
• Whether more groups get access through the Fairwind program.
• Whether independent tests confirm Google's claims about Carbon's skills and safety.