
Gemini 4 Carbon Went to Cyber Defenders First. That Order Matters.
Google's newest frontier model is, by Google's own benchmarks, very good at finding software flaws. So the company released it to vetted defenders first and the public two weeks later. Here is what the numbers say and why the rollout order is the real story.
When a model gets good enough at finding security holes, the question of who gets it first stops being a marketing decision. Google appears to agree. On October 28, 2026, it announced Gemini 4 Carbon, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities," and then gave it only to a subset of cyber defenders through what it calls the Fairwind Program. Developers, businesses and Google AI Ultra subscribers got access on November 10.
Two weeks is not long. But it is a deliberate head start for the people who patch, and it is the first time we have seen a major lab sequence a general release around cyber risk this openly.
What Carbon Can Do, in Google's Numbers
Google's launch post lists three external cybersecurity benchmarks, all "evaluated without guardrails." Benchmarks are lab conditions, not the real internet, and these are Google's reported scores, not independent audits. Still, they describe the shape of the capability.
| Defense stage | Benchmark | Carbon score (Google) |
|---|---|---|
| Finding real vulnerabilities | SEC-bench Pro | 88.2% |
| Proving a flaw is exploitable | ExploitGym | 54.8% |
| Patching the root cause | CWE-bench | 71.0% |
Google also says Carbon ranked first on version 5.0 of the Artificial Analysis Intelligence Index at 59, and that internally it showed "more than 2x uplift in success rate for generating secure code" over Gemini 4 Argon, the model Google released earlier in October.
What Defenders Did With It in Two Weeks
The November 10 availability post gives three early examples. We report them as Google described them; outside verification is not yet public.
- Project Fortify: on Google's internal benchmark, Carbon hit "100% recall on scanning and 95% precision on triage" in Google's autonomous scanning and remediation system.
- Cloud Vulnerability Research: Google's team used Carbon to prove that two bugs in "a popular open-source software" were exploitable, helping maintainers re-classify the bug as critical and ship a patch. Details are being withheld until responsible disclosure finishes.
- Wiz Scan for Good: when scanning two major SaaS providers, Google says the model "uncovered 0-day vulnerabilities in under two hours," and the team worked with the affected companies to fix them.
Field note: Notice what is withheld. Google names no vulnerable project and publishes no exploit details while fixes roll out. That is exactly the white-hat norm: disclose to the people who can patch, then tell everyone else.
The Same Skill Cuts Both Ways
Here is the uncomfortable part. A model that scores 54.8% on an exploitation benchmark is, by definition, useful for exploitation. Google's own threat team published a report on November 16 describing criminals already chaining language models into automated attack pipelines. The capability is dual-use, and the company's launch post says so in its own way: it "strengthened our misuse safeguards to help ensure stronger defensive reasoning does not translate into offensive risk."
Google reports several guardrail results for the public version:
- In simulated attack environments, Carbon "blocked 99% of attack trajectories" from an adaptive red-teaming agent, with a per-attempt breach rate of 0.8%, down from 3.1% on Argon.
- In Gray Swan's testing, the automated cyberattack query success rate fell to 2.17%, compared with 9.86% on Argon.
- Trajectory Labs "found zero new universal jailbreaks across 132 hours and over 13,000 requests."
- On Gray Swan's indirect prompt injection benchmark, Google reports a 0% attack success rate, while adding that "prompt injection robustness remains an area for continued improvement."
The Messy Middle: Two Models, Two Access Levels
The fine print matters. Google describes "Gemini 4 Carbon for cyber," accessed through the Fairwind Program, as the version that "leads on key benchmarks across finding, exploiting and patching." The public model carries the extra safeguards described above. So defenders and the general public are not using exactly the same thing, which is the point of the program, and also why benchmark scores in headlines should be read carefully.
Pricing is public. Google lists an introductory rate of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30, 2026, for the public Carbon model.
Decision Matrix: What This Changes for You
- If you lead a security team: apply for defender access programs from every major lab. Two weeks of lead time is worth having.
- If you maintain code: run AI-assisted scanning on your own repositories before someone else does.
- If you are a small business: the practical takeaway is speed. Keep automatic updates on and ask your web host how fast it patches.
- If you set AI policy at work: separate "can read our code" from "can run commands." Agentic tools need the second permission much less often than people assume.
The Inflection Point
For most of software history, the people best at finding flaws were rare and expensive, and they chose sides one at a time. Models like Carbon turn that skill into a service. When the skill is a product, the most important safety decision is distribution: who gets the strongest version, when, and with what checks. Google's answer this time was "defenders first, with guardrails for everyone else." It is a small head start, but it sets a precedent we hope every lab follows, and one that regulators will likely start asking about.
Sources
We checked every fact, figure and quote in this story against the sources below. Quotes are reproduced as published. Images are AI-generated illustrations, not photos of the events described.


