Atlas AI Resources

← Back to AI News

AI models

Google’s new Gemini 4 Carbon model goes to cyber defenders first. Here is why.

Published October 30, 2026

On October 28, Google announced Gemini 4 Carbon, which it calls its strongest AI model. Google says Carbon is especially good at finding, testing and fixing security holes in software, so for now only a small group of vetted security teams can use it. Developers, businesses and regular users get it later, once Google finishes testing its safety limits. An independent tester ranks it first overall on its main scorecard, though not first at everything.

What happened

Carbon is a frontier model, meaning one of the most capable AI systems a lab has built. It follows Gemini 4 Argon, which Google launched earlier this month. In its announcement, Google DeepMind’s Koray Kavukcuoglu says Carbon is built for long, multi-step work, and that Google’s own teams have seen it run for days on hard engineering problems such as tracking down subtle bugs in large codebases.

The security claims are the headline. A vulnerability is a flaw that lets an attacker misuse software, an exploit is working code that takes advantage of one, and a patch is the fix. Google reports that Carbon scored 88.2% on SEC-bench Pro (finding real flaws), 54.8% on ExploitGym (turning flaws into exploits) and 71.0% on CWE-bench (writing working fixes). Each is a benchmark, a standard test used to compare models. Google says it ran these tests with the model’s guardrails, its built-in safety limits, switched off, and that Carbon beat models from OpenAI and Anthropic in its comparison.

Access starts this week for a subset of cyber defenders through Google’s Fairwind Program, which Google launched in September for governments, critical infrastructure operators such as energy and healthcare networks, and major tech platforms. Google says Fairwind has more than 650 partners, and that they must limit access to their own security teams and use protections like multi-factor login. Google also says it is taking part in the US government’s voluntary pre-release testing of AI models. Pricing starts at an introductory $2 per million input tokens and $10 per million output tokens (a token is a small chunk of text, roughly part of a word), rising to $4 and $20 later.

Why it matters to a normal person

Almost everything you use, from your bank’s app to your hospital’s records, runs on software with flaws nobody has found yet. A model that finds and fixes them quickly helps defenders. The same skill could help attackers, and that is why Google is letting defenders go first: the idea is to give them time to patch before similar tools spread. We covered a real example of AI-found flaws this week, in PHP.

Google also claims Carbon behaves better. It says Carbon makes about half as many unsupported claims as Argon, that high-severity misbehavior alerts in its internal coding work dropped by almost 80%, and that it resisted every attempt in one prompt injection test, where hidden instructions in a web page or file try to hijack the AI.

What it does not mean

It does not mean you can use it. Carbon is not public yet, and Google has given no date beyond “as soon as possible.”

It does not mean the security and safety numbers are proven. They are Google’s own results, and several come from internal tests outsiders cannot check. Scoring 88.2% on a test of known flaws is not the same as finding 88% of the flaws in real software.

It does not mean Carbon is best at everything. Artificial Analysis, a company that runs its own tests on AI models, gives Carbon the top overall score of 59 on its index, one point ahead of Anthropic’s Claude Opus 5.5. But it places Carbon second on its coding-agent ranking, behind Claude Opus 5.5, and behind several rivals on office-style tasks. It also found Carbon’s hallucination rate, how often it gives a wrong answer instead of admitting it does not know, rose from 4% for Argon to 11% on one test, though that is still among the lowest it measured.

Related: AI helped find two old flaws in PHP, which runs much of the web. The US now lists them as exploited., Google opened Gemini 4 Argon to paying users. Free access is still missing

Sources