Google Gemini 4 Carbon: the AI built to find and fix security holes, and why only defenders get it first
Google's new Gemini 4 Carbon tops a major AI ranking and sets records at finding, exploiting and patching software vulnerabilities. Because those skills could be misused, Google is giving it to vetted cyber defenders first.
Google has released Gemini 4 Carbon, which it calls its “strongest model”, with record-breaking skills in cybersecurity. But you can’t use it yet. Google is giving it first to vetted cyber defenders, because the same abilities that help protect software could also help attackers.
Carbon was announced on October 28, 2026, two weeks after Gemini 4 Argon.
How good is it?
- #1 overall: it ranks first on the Artificial Analysis Intelligence Index (v5.0), a popular benchmark, with a score of 59.
- Finding vulnerabilities: 88.2% on SEC-bench Pro
- Exploiting vulnerabilities: 54.8% on ExploitGym
- Fixing vulnerabilities: 71.0% on CWE-bench
- Secure code: Google says it writes secure code more than twice as successfully as Argon.
Google’s Chrome team reported that Carbon was more than 10 times better than another lab’s large model at creating high-quality security patches. Security firm Wiz says it leads on web penetration-testing tests.
It can work for days
Google says its teams have seen Carbon work on difficult engineering tasks for days, such as tracking down and fixing bugs in huge codebases, working through dead ends on its own.
Why can’t everyone use it?
An AI that’s great at finding and exploiting security holes is a double-edged sword. In the right hands, it helps companies fix weaknesses before criminals find them. In the wrong hands, it could help attackers. So Google is:
- Giving access first to a small group of cyber defenders
- Expanding gradually to trusted partners through its Fairwind Program
- Taking part in the US government’s voluntary pre-release model testing
Broader access for developers, businesses and consumers is planned “as soon as possible,” but Google hasn’t given a date.
We’ve reached the point where AI companies are deciding who gets their best models first, based on how they might be used.
Safety improvements
Google says Carbon is also safer than Argon:
- Prompt-injection attacks (tricking the AI with hidden instructions) succeeded 0.0% of the time in one major test
- High-severity misbehavior alerts in internal coding tests fell by almost 80%
- Unsupported or made-up claims were cut by about half
Pricing (for approved users)
- Introductory: $2 per million input tokens and $10 per million output tokens
- Later: $4 and $20 per million tokens
Why it matters
This month has been full of security incidents: critical PHP flaws, the Railway breach, the ConnectWise hack and attacks on AI coding tools. AI models like Carbon could help defenders keep up, but they also show why careful release matters. Read more: the week the internet had a security crisis.
Frequently asked questions
What is Gemini 4 Carbon?
Google’s newest and strongest AI model, released October 28, 2026, with record cybersecurity abilities.
Can I use Gemini 4 Carbon?
Not yet. It’s limited to vetted cyber defenders for now, with wider access planned later.
Why is Gemini 4 Carbon restricted?
Because its skills at finding and exploiting vulnerabilities could be misused by attackers, Google is releasing it to defenders first.
What’s the difference between Gemini 4 Argon and Carbon?
Carbon is the newer, stronger model, especially for cybersecurity and long tasks, and Google says it’s safer too.
