# Gemini 4 Carbon improves cybersecurity and Google is not opening it up

> Google unveils Gemini 4 Carbon, a leader in detecting and patching flaws, and reserves it for defenders in the Fairwind program with no date for general release.
> By José Maldonado · Artificial Intelligence · 2026-10-31

[Google presented Gemini 4 Carbon on October 28](https://blog.google/technology/google-deepmind/gemini-4-carbon/), a frontier model that the company places ahead of its competitors in three specific tasks: finding vulnerabilities in real software, turning them into exploits and writing the patches that close them. The figures it publishes are 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear's patching benchmark. These are numbers that, in any other context, would serve to sell a subscription to everyone.

The problem is that Google is not selling it. Carbon is being delivered this week to a subset of cybersecurity defenders and will be expanded to trusted partners through the Fairwind program, while the company gathers feedback and adjusts its safety barriers. The company says it will open it to developers, businesses and consumers "as soon as possible", without giving a date, and does not confirm that the API is available today.

That contrast between capability and access is what turns the announcement into a story with substance. A model that finds flaws better than a human team and writes the exploit in the same session is not just another product in the model race: it is a dual-use tool that governments want to oversee before it circulates. Google has understood this and has decided to show the muscle without opening the door.

## The numbers Google puts on the table

The three tests the company cites measure different things and it is best not to conflate them. SEC-bench Pro evaluates the ability to locate vulnerabilities in real projects, ExploitGym measures whether the model manages to exploit them and [Collinear's bank scores the quality of the patch](https://cwe-bench.com/) that it proposes. Winning on all three at once means covering the full cycle, from discovery to fix.

Model

Patching (Pass@1)

Patching (Pass@4)

Cost per run

Gemini 4 Carbon

71 %

81 %

Not published

Gemini 4 Argon

68 %

75 %

$6.63

GPT-6 Astra

68 %

74 %

$2.85

Claude Opus 5.5

67 %

79 %

$0.79

Claude Fable 5.1

58 %

73 %

$3.47

Collinear's bank table adds a nuance that Google does not highlight: the cost per attempt for Carbon is not listed, while its competitors range from under a dollar to almost seven. If the model is much more expensive per task, the performance advantage may evaporate in industrial use, where thousands of runs are launched to find a flaw.

It is also worth remembering who commissions the tests. All three results are third-party measurements or internal comparisons, and the competitor comparison section is signed by Google. This is standard practice in the industry and does not invalidate the data, but it does argue for reading them as a snapshot from the vendor rather than an independent ranking.

## What it means to win in search and exploitation

Finding vulnerabilities is the part that matters most to a defending company. A model that reviews code at scale and flags exploitable paths reduces the time between the publication of a flaw and its detection, and that time is what separates a contained incident from a breach. Google claims that Carbon generates secure code with more than double the success rate of Argon in its internal tests.

Exploitation is the uncomfortable part. A model that turns a known vulnerability into a working exploit serves both the red team and the attacker, and the difference between the two is not in the model. The same capability that lets a defender verify that a patch works lets a criminal have a weapon on the same day the flaw is published.

This dual use has already been seen this month with CVE-2026-98432, the Linux kernel flaw that an Anthropic research model turned into an exploit with a root shell while Ubuntu and Red Hat remained unpatched. When the defensive community talks about compressing the exposure window, this is exactly what it means: AI shortens the time in both directions.

## Access is the product

The decision to limit Carbon to the Fairwind program is itself the news. Google describes a phased delivery, first to selected defenders and then to trusted partners, with the aim of gathering feedback and adjusting barriers before opening it up. The company adds that it is taking part in the US government's voluntary pre-release testing process.

That detail places Carbon in a different category from a commercial model. Governments want to see a tool capable of finding and exploiting flaws before it is published, and the companies that develop it need to maintain a smooth relationship with those governments to avoid restrictions. The result is a product that exists, that works and that almost no one can buy.

For the market, the consequence is a strange mix of messages. The model race is measured in scores, but competitive advantage is measured in access, and access depends on a program curated by the vendor. Anyone who needs that capability and is not on the list will be left with competitors' models and with the question of how much they are losing.

## The week AI was sold as a public service

Carbon does not arrive alone. In recent days, [OpenAI has moved its mid-tier model toward lower prices](https://observatorioblockchain.com/en/ai/gpt-6-1-terra-mitad-precio-uso-pc/) and Anthropic has presented its stock market listing, while Google has racked up two frontier launches in a month. Competition has shifted from "who has the smartest model" to "who controls how it is used," and all three companies have understood that the corporate buyer wants guarantees before demonstrations.

That shift explains why Carbon presents itself with a national defense pitch rather than a productivity one. If the argument is that the model protects critical infrastructure, restricted access justifies itself, and the company places itself on the side of the good guys without giving up the capability. It is good communication and also a way of dodging the question about offensive use.

The question left unanswered is what happens when the same capability reaches an open model. Carbon's weights will not be published, but those of other models will, and an open model with half the capability and no access restrictions can be more dangerous than the best closed model in the world. The barrier to entry for writing exploits is no longer knowledge, it is the subscription.

## A model Google is not putting up for sale

Carbon leaves three data points worth holding on to. The first is that AI now covers the full lifecycle of a vulnerability, from discovery to patch. The second is that access has become the strategic variable, more than the score. The third is that the price of that capability is still unclear, and without a price there is no market.

For a security team, the practical reading is that the in-house model is not the lever. The lever is the process: software inventory, fast patches and logs that make it possible to know what happened. Carbon speeds up those who already have that process in place and does not fix those who do not. For the rest of us, the news is that the best tool for finding flaws exists and is not in the store.
