Gemini 4 Carbon: Google built a model that hunts for holes

Google released Gemini 4 Carbon — a model with one job: security. Not "smarter in chat", but "finds the hole, proves it is live, writes the fix". Two weeks ago the company widened access to Gemini 4 Argon for paying users and developers. Now comes the version you cannot buy off the shelf.

Three tests, three different skills

Google reports Carbon ahead in three checks — vulnerability discovery, exploit verification and remediation. The order matters.

Finding a hole is half the job. Proving it actually fires is a second skill. Shipping a patch that does not break the code next door is a third. There are plenty of scanners on the market; decent auto-fixes are rare. A model that covers all three stages hits the most expensive part of a pentester's day: not the search, but the write-up and the repair.

Who gets in, and when

iThome reports that a slice of defenders gets access first, with the circle widening later. TechNews repeats Google's framing: defense first. In practice this is the standard playbook for shipping a "dangerous" model — a showcase for enterprise buyers, not for consumers.

Which is the point: Google is not selling a subscription, it is selling clearance. In enterprise security, access to the tool is the product.

Where the nerves start

A model that verifies exploits can also write them. Limiting access to defenders is the sane call. It holds exactly until the first weight leak or the first rival that opens an API "for everyone".

While OpenAI and Anthropic argue about who is more responsible, Google ships a model that monetises safety directly. That is a quiet shift: AI moves out of chat windows and into infrastructure, where a mistake costs more than a bad answer.

Another detail: Carbon lands two weeks after Argon. That pace says Google runs several lines at once, each tuned to a task. The race for one giant model to do everything is quietly ending.

What to watch

First, false-positive numbers. A model that finds a lot and is wrong a lot is worse than an ordinary scanner. Second, how fast its patches clear human review. Third, whether rivals launch their own "safe" editions — enterprise defence pays steadily, with no seasonality.

For Asian vendors this is a signal too. Taiwan and Korea sit on infrastructure attacked daily, and they buy tools early. If Carbon lives up to the claims, the next perimeter-defence contract will not go to a local startup.

Prediction: Within six months, at least one major AI vendor will launch its own restricted-access "security" line, following Carbon's template.

Disclaimer: not officially confirmed — Google's blog is not in our source registry; facts come from two independent Taiwanese outlets.