Google’s Gemini 4 Carbon: A New AI Model Built to Find Security Flaws
Google has introduced its most powerful AI model yet, and its first users will not be ordinary consumers. Gemini 4 Carbon is designed to find and fix weaknesses in software, the kind of holes hackers use to break into systems. For now, Google is handing it only to a small group of vetted cyber defenders while it tests safeguards. The launch comes just two weeks after Gemini 4 Argon and marks a shift toward AI built for high-stakes work. Here, we explain what this AI cybersecurity model does, how it performed in Google’s tests, and what remains unproven.
What Google Announced
Google unveiled Gemini 4 Carbon on October 28, describing it in its official blog post as its strongest model, with leading cybersecurity and long-horizon capabilities. Long-horizon means the model can keep working through extended, multistep tasks without losing track of the goal.
Google says it also improved Carbon for full-stack web development, scientific work and understanding images, building on feedback from the Argon release.
How It Helps Security Teams
Cyber defense usually follows a cycle: find a weakness, confirm it can actually be exploited, then fix it properly. Google designed Carbon to help at each stage.
- Finding flaws. The model searches real software for vulnerabilities that attackers could exploit.
- Checking risk. It tests whether a suspected flaw is genuinely dangerous, helping teams prioritize.
- Fixing the root cause. Rather than patching symptoms, it aims to repair the underlying problem in the code.
For overstretched security teams, an assistant that handles these tasks could speed up work that normally takes skilled people days or weeks.
The Benchmark Claims
In Google’s testing, Carbon scored 88.2% when hunting for flaws on SEC-bench Pro, 54.8% when exploiting them on ExploitGym and 71% when repairing them on CWE-bench. It also sits at the top of the latest edition of the Artificial Analysis rankings, with a score of 59, according to Technology.org.
Google also says Carbon follows instructions more reliably, cutting cases where it overstepped its task by 26% and cases where it fell short by 29% during multistep work.
Why caution is warranted
These numbers come from Google, and outside experts cannot yet test the model, because there is no public access. Some reporting also notes that Google’s security tests reportedly removed the normal guardrails, so the model faced looser limits than it would in daily use, as MindStudio observed. Independent evaluation will matter once the model opens up.
Who Gets Access First
Access starts with a hand-picked group of cyber defenders this week, expanding through Google’s Fairwind Program before developers, businesses and consumers get it. Google has not given a date for wider release.
For developers, Google has set a launch price of $2 for every million input tokens and $10 for every million output tokens, matching Argon’s starting rate.
Why Limited Release Matters
An AI that excels at finding software flaws is valuable to defenders, but the same skill could help attackers if misused. A staged rollout gives Google time to monitor how the model is used and to strengthen its protections. This tension sits at the heart of modern AI security. Recently, AI tools helped uncover serious flaws in widely used web software, as we reported in our coverage of the PHP2Shell vulnerabilities.
Where Carbon Fits in Google’s AI Push
Carbon arrives during an intense stretch for Google. The company recently launched Gemini 4 Argon and the interactive world model Genie 4, while its latest earnings showed how heavily it is investing in AI infrastructure. Our AI security tools for businesses guide covers wider options.
A Defender’s Tool, Released With Care
Gemini 4 Carbon shows AI moving into one of technology’s most sensitive jobs: protecting software from attack. Its claimed abilities are impressive, but they remain unverified until outsiders can test them. Watch for wider access and independent results before judging how much Carbon really changes cybersecurity.
Frequently Asked Questions
Not yet. Access is limited to selected cybersecurity defenders, with broader availability planned later.
Carbon is Google’s newer, stronger model, with a particular focus on cybersecurity and long multistep tasks.
Google is charging $2 for each million input tokens and $10 for each million output tokens at launch.
No. They come from Google’s own testing, and independent checks are not yet possible.
