Google's Gemini 4 Carbon is built for cybersecurity, and starts with a small group
Google's new top model is aimed at finding and fixing software flaws, but only a few security professionals can use it for now.
The short version
- Gemini 4 Carbon targets cybersecurity and tasks that run over several days
- Google says it more than doubles the success rate of its predecessor at writing secure code
- It ranks first on the Artificial Analysis intelligence index, per the report
- This week it goes only to a small group of security defenders
- Wider access will come in stages through Google's Fairwind program
Why it matters
A model that is good at finding and patching software flaws helps defenders, but the same skill could help attackers. That is why Google is limiting who gets it first. For ordinary users, the practical effect should be safer software over time, not a new app to download today.
What Google released
Google has launched Gemini 4 Carbon, a new frontier model, meaning one at the leading edge of what AI can do. As Sina Finance reported, Google built it around two goals: cybersecurity and long, complicated jobs.
The report says it tops the Artificial Analysis intelligence index, a third-party ranking. It also placed fourth on WebDev Arena, a leaderboard for building web apps, and improved on its predecessor there.
A model that hunts for software flaws
Google says Carbon is meant for work that runs across several days, such as tracing the root cause of a deep system failure in a large codebase. It is designed to drop dead ends on its own and keep going.
Security is the main upgrade. The model is described as strong at finding vulnerabilities (weak spots in software), checking they are real, and writing patches to fix them. In Google's internal tests, it more than doubled the previous model's success rate at producing secure code. Google's Chrome team also compared patch quality against other labs' models and said Carbon came out well ahead.
Safer behavior, by Google's measures
The report lists several behavior improvements, all from Google's own testing. Made-up answers fell by about half. When evidence is thin, the model is meant to say what it cannot do and what information it lacks.
- It scored zero on successful cheating in one coding test.
- It resisted every indirect prompt-injection attack in the tests. That is when hidden instructions in a document or web page try to hijack an AI.
- It is less likely to overstep or under-deliver on a task.
These are company-reported results, not independent checks.
Why access is limited
Google is not opening the model to everyone. This week it goes to a small number of cybersecurity defenders, in step with a voluntary US government process for testing models before release.
After that, Google plans to widen access gradually to trusted partners through its Fairwind program. Developers, companies and the general public come last, once feedback has been gathered and the safeguards improved. Pricing starts the same as Gemini 4 Argon, with a promotional period followed by higher standard rates.
What to watch
- When Google moves beyond security professionals to developers and the public
- Whether independent testers confirm the security claims
- How Google keeps attackers from misusing a strong security model
Try it
Turn on two-step sign-in for your email and main accounts today. It blocks many attacks, however good the tools on either side become.
Sources and links
- ReportingSina Finance: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Sina Finance · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
