Gemini 4 Carbon Is Here: What an AI Built for Cyber Defense Means for Teams

Google gave Gemini 4 Carbon to cyber defenders first and opened it to all paying developers on 10 November. What its security numbers show, and how smaller teams can use it safely.

Published 6 min read

On 10 November 2026, Google made Gemini 4 Carbon generally available to every paying developer and enterprise on its API. Two weeks earlier, the same model was only open to a hand-picked group of cyber defenders. That sequence says a lot about where frontier AI is heading: the most capable models are now judged not only by how well they write code, but by how well they can find, exploit and fix security flaws in it.

This article looks at what Google actually released, what its numbers do and do not show, and what a small or mid-sized team should do with a model like this.

What Google released, and in what order

Google announced Carbon on 28 October 2026 as its "next frontier model". It did not open it to everyone. It started with "a subset of cyber defenders", with access widening to trusted partners in its Fairwind Program.

Fairwind itself is older. Google launched it on 2 September 2026 as a limited-access programme for governments and trusted partners, starting with an earlier cyber-tuned model and its CodeMender tool. Google says more than 650 partners take part, and they must limit access to their internal security, incident-response or penetration-testing teams and use multi-factor authentication.

Then, on 10 November, Google opened Carbon to all paying developers and enterprises through Google AI Studio and Gemini for business, and began rolling it out to Google AI Ultra subscribers in its consumer apps. The Gemini API release notes list the same date for gemini-4-carbon becoming generally available.

Date

Step

2 September 2026

Fairwind Program launches for governments and trusted partners

28 October 2026

Carbon announced; access for a subset of cyber defenders

10 November 2026

General availability for paying developers and enterprises

30 November 2026

Introductory price ends

Why a "cyber" model is released differently

Most model launches are a single day: announcement, API, pricing. Carbon was staged because the skills Google highlights are dual-use. A model that is good at finding a vulnerability in your code is, by definition, good at finding one in someone else's.

Giving defenders a head start is a deliberate bet. If security teams can scan and patch their systems before the same capability is widely available, the balance tips slightly toward defence. Google also says it took part in "the U.S. government's voluntary process for pre-release model testing" before the launch.

The same tension shows up across the industry. In August, Help Net Security reported on an incident in which, in its words, an "agentic collective" autonomously broke into OpenAI's research infrastructure by chaining previously unknown vulnerabilities with credentials leaked online. According to the same report, OpenAI's response leans on AI as well: tools that validate code changes, AI triage of almost all initial security alerts, and models that search for attack paths such as configuration errors and excessive permissions.

The pattern is clear. Attackers and defenders are getting the same class of tools, and what differs is who uses them first and how carefully.

What the numbers show, and what they do not

Google publishes several benchmark results for Carbon's security skills. In the launch post they cover the three stages of the work:

Task

Benchmark

Carbon's score (Google's figure)

Finding vulnerabilities

SEC-bench Pro

88.2%

Exploiting them

ExploitGym

54.8%

Patching them

CWE-bench

71.0%

Three points of caution are worth keeping in mind:

  • These are vendor numbers. Google ran or reported them, and it notes they were "evaluated without guardrails". Independent results will take time to appear.

  • Benchmarks are not your codebase. A score on a curated set of known bugs says little about a 10-year-old internal system with no tests and odd dependencies.

  • Exploitation scores cut both ways. A model that succeeds in more than half of exploitation tasks is useful for proving a bug is real, and also the reason access was staged.

Google also reports safety figures for the general release. In its simulated attack tests, it says Carbon "blocked 99% of attack trajectories", with a per-attempt breach rate of 0.8%, down from 3.1% for Gemini 4 Argon. It also cites external testing by Trajectory Labs that found no new universal jailbreaks across 132 hours and more than 13,000 requests. These figures matter, but they describe the safety layers around the model, not a guarantee that misuse is impossible.

Where it can help a smaller team

Most companies do not have a red team. They have one or two engineers who also handle security, a backlog of dependency warnings and a vague sense that something is out of date. That is where a strong security-aware model can genuinely help:

  • Reviewing code before it ships. Ask the model to review a change for injection, authentication and permission mistakes, and to explain each finding with the exact lines involved.

  • Turning scanner noise into a short list. Automated scanners produce long reports. A model can group duplicates, explain which issues are reachable from the outside and propose a fix order.

  • Drafting patches and tests. For a confirmed bug, it can draft both the fix and a test that would have caught it. A developer still reviews and merges.

  • Explaining incidents. Given logs from your own systems, it can help reconstruct what happened and in what order, which is often the slowest part of a response.

The common thread is that the model works on your own code and systems, and a person decides what to do with the result.

How to use it without creating new risk

A security model that can read your code and run tools is itself a powerful agent. The usual rules for agents apply, and a few more:

  1. Only point it at systems you own or are authorised to test. Scanning someone else's infrastructure is not a grey area, whatever tool does it.

  2. Run it in a sandbox. Give it a copy of the repository and a test environment, not production credentials.

  3. Keep the permissions narrow. Read access for review; write access only to a branch, never directly to the main line.

  4. Require a human before anything changes. Patches, configuration changes and disclosures should wait for approval. We covered how to design those checkpoints in When should an AI agent ask before it acts?.

  5. Check what you send. Source code, logs and customer data sent to an external API leave your network. Review the provider's data terms and strip secrets before you start.

  6. Verify findings. Treat every reported vulnerability as a hypothesis until someone reproduces it.

What to watch next

The introductory price is $2 per million input tokens and $10 per million output tokens. Google says that after 30 November 2026 the price rises to $4 and $20 for public access, so teams planning large scans should budget for the higher rate.

The more important signal will be independent evaluation. Over the coming months, look for results from security researchers who were not part of the launch, and for reports from Fairwind partners on what the model found in real systems. Those will say more than any launch table.

For teams in our region, the practical advice is the same whichever model you use: start with your own code, keep a human in the loop and measure what the tool actually finds. More guides on working with AI safely are on the Hamroshd blog.

Sources