Google tests its Gemini 4 Carbon AI model with the US government

Google’s newest AI model, Carbon, is reportedly being tested with the US government. The pitch is simple: an AI that is very good at finding software flaws should probably be pointed at the good guys’ problems first.

The model’s full name is Gemini 4 Carbon. Alphabet announced it on October 28, 2026, as a frontier model built to sharpen cybersecurity work, from spotting vulnerabilities to exploiting and patching them.

What Carbon actually does

The benchmark numbers are where Carbon makes its case. On SEC-bench Pro, which tests the ability to identify real-world vulnerabilities, the model hit an 88.2% success rate.

Exploitation is a tougher assignment. On ExploitGym, a benchmark for exploitation tasks, Carbon scored 54.8%.

Patching held up well too. The model recorded a 71.0% success rate on Collinear’s CWE-bench, which measures how well an AI fixes known categories of software weakness.

Then there is secure code generation. Carbon reportedly doubled the success rate of its predecessor, Gemini 4 Argon, at writing code that does not ship with its own security holes.

Prompt injection got a clean sheet as well. That is the trick where an attacker hides malicious instructions inside content an AI reads, hoping the model obeys them. On the Gray Swan benchmark, such attacks against Carbon had a 0.0% success rate.

A crowded release calendar

Carbon did not arrive in a vacuum. Gemini 4 Argon launched just two weeks before it, which says something about the pace inside Google’s Gemini program.

On the broader leaderboard, Carbon posted a score of 59 on the updated Artificial Analysis Intelligence Index v5.0. That result puts it at number one in its category, ahead of earlier Gemini models and rival systems on key cybersecurity benchmarks, according to the research findings.

Access remains tight. Carbon is currently available only to a selected group of cybersecurity defenders through what Google calls the Fairwind Program, with initial access beginning this week as part of a phased rollout.

The US government testing sits alongside that limited release.

Wall Street took note, though it did not exactly throw a parade. Alphabet shares rose about 1.5% after the announcement.

Background: why cyber is the new AI battleground

The research findings place Carbon in direct competition with OpenAI and Anthropic in coding and cyber defense, where each lab is trying to prove its models are both capable and safe to deploy.

What this means

For Alphabet investors, Carbon adds another data point to the argument that Google can compete at the frontier. The roughly 1.5% share bump signals that the market sees specialized, high-value models as a meaningful part of the company’s AI story.

The bigger question is monetization. Cybersecurity is a market where buyers pay for results, and benchmarks like an 88.2% vulnerability discovery rate are the kind of figures that can show up in enterprise sales decks.

A model that scores 54.8% on exploitation tasks is, by design, useful to attackers as well as defenders. How carefully Google manages the move from a selected group of defenders to wider availability will shape how regulators and customers view the product.

For competitors, Carbon raises the bar in a specific and measurable way. Rivals will likely face pressure to publish comparable cybersecurity benchmarks, particularly on prompt-injection resistance, where a 0.0% attack success rate sets a hard target to match.