Google's new security model is not for you yet, and that is the point
The short answer
Gemini 4 Carbon is a model built for finding and fixing security flaws, and Google has released it only to a small group of trusted cyber defenders, with the US government testing it. Nobody can sign up. The scores are strong, 88.2 percent on a benchmark of real vulnerability finding, but the release pattern tells you more than the score: Google is handing a capable tool to defenders first because the same capability works for attackers.
What it is
Google DeepMind announced Gemini 4 Carbon on 29 October, two weeks after Gemini 4 Argon. Where Argon is a general frontier model, Carbon is aimed at one job: read code and systems, find the flaws, and write a working fix. Google says it scores 88.2 percent on SEC-bench Pro, which tests whether a model can find real software vulnerabilities, and 71 percent on CWE-bench, which tests whether it can produce fixes for whole categories of weakness. It also topped the Artificial Analysis Intelligence Index at 59, according to Google's own announcement.
Who gets it
A subset of trusted cyber defenders first, and the US government is testing it. That is a deliberate choice, not a delay. Every serious AI lab now treats exploit-capable models the way earlier generations treated biology models: useful enough that you cannot keep them locked up forever, dangerous enough that the first users are chosen by hand.
The context matters. This month security researchers have been calling Black October after a wave of AI-linked incidents, and in the same week a Chinese model was reported having completed three of thirteen full attack chains in a simulation. Capability on the defensive side and capability on the offensive side are not separate stories.
Why it matters to a normal business
You will not get access this quarter. What you will get, eventually, is cheaper and faster vulnerability scanning inside the tools you already use, most likely as a feature rather than a product. Until then the practical advice does not change: patch what your scanner tells you, and do not assume a model finding flaws means your code is safe.
If you run agents with access to your systems, the permission question from our agent access guide is still the one that decides your risk, not the benchmark score.
Questions people are asking
Can I use Gemini 4 Carbon?
Not yet, unless you are in the selected group of cyber defenders or part of the US government testing. There is no public signup.
How is it different from Gemini 4 Argon?
Argon is a general frontier model released two weeks earlier. Carbon is specialised for finding vulnerabilities and writing fixes.
What do the benchmark numbers mean?
SEC-bench Pro tests whether a model can find real software vulnerabilities, and CWE-bench tests whether it can produce working fixes for categories of weakness. Higher is better, and Google published the figures itself.
Does this make my website safer?
Not directly, and not now. It may show up later as better scanning inside tools you already pay for.
Why does Google restrict access?
Because a model that can find and exploit flaws is useful to attackers too. Restricted release keeps the first users on the defensive side.
Sources
Google DeepMind announcement, 29 October 2026 · The Neuron daily digest, 29 October 2026 · Reuters reporting on AI security incidents, 29 October 2026.