Google company sign on Charleston Road in Mountain View, CaliforniaFile photo: The Google company sign on Charleston Road in Mountain View, California, in 2022. Photo: Dietmar Rabich via Wikimedia Commons (CC BY-SA 4.0)

Google says its new Gemini 4 Carbon model is the strongest AI system it has built, and the best yet at finding, exploiting and fixing software flaws. That is exactly why the public cannot use it. For now, only a vetted group of cyber defenders gets access.

By HahuPrime News Desk | October 29, 2026

MOUNTAIN VIEW, California — Inside Google, engineers have watched their newest AI model work on a single problem for days. It hit dead ends, backed up and kept going until it had traced a subtle bug through a large codebase and written a fix.

That, at least, is how Google describes Gemini 4 Carbon, the frontier model it announced on Wednesday. The company calls it its strongest model yet. It also says almost nobody outside the company will be allowed to touch it for the time being.

Carbon is going first to “a subset of cyber defenders” through Google’s Fairwind Program, wrote Koray Kavukcuoglu, a senior vice president at Google DeepMind and the company’s chief AI architect. Access will widen in phases to more trusted partners. Developers, businesses and ordinary users will get it “as soon as possible,” with no date given.

The reason is the same skill that makes the model useful. A system that can find a hidden hole in software can help a bank or a hospital patch it. In the wrong hands, it can help someone break in.

Top of the table, by Google’s count

Google says Carbon now ranks first on version 5.0 of the Artificial Analysis Intelligence Index, a widely watched combined benchmark, with a score of 59. The Verge reported that it sits ahead of OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 on that index.

The security numbers are the headline. On SEC-bench Pro, a test of whether an AI can uncover real vulnerabilities in real software, Google says Carbon scored 88.2 percent. On ExploitGym, which measures whether a model can turn a flaw into a working attack, it scored 54.8 percent. On CWE-bench, which checks whether a model can write a fix that addresses the root cause, it reached 71 percent. Google says all three external tests were run without its guardrails switched on.

The company also cited its own teams. “The Chrome team has reported that Carbon performed more than 10X better than another lab’s large frontier model when generating high quality patches,” Kavukcuoglu wrote. On internal tests, he said, Carbon was more than twice as successful as its predecessor at producing secure code.

Those claims come with a caveat. Every figure so far is Google’s own, and none has been independently checked. Ars Technica, which reviewed the benchmark charts, noted that Carbon beats Gemini 4 Argon, released only two weeks earlier, “by slim margins in every test,” and that the gap between Carbon and its rivals “is just a few points.” The site raised the possibility that Carbon could be a post-training upgrade of Argon rather than an entirely new model. Google presents it as a new model.

Google also pointed to safety results. On Gray Swan’s indirect prompt-injection benchmark, which tries to trick an AI into following hidden instructions planted in web pages or documents, Carbon resisted every evaluated attack, according to the company. Argon had a 0.7 percent attack success rate. It also said Carbon cut “unsupported claims” roughly in half compared with Argon.

Who gets in first

Fairwind is Google’s gatekeeping programme for its most capable security models. Its public page says it works with more than 650 partners worldwide. Priority goes to governments and national cyber authorities, operators of critical infrastructure such as healthcare, telecommunications, energy and banking, and the core technology platforms that millions of other systems depend on.

The rules are strict. Google runs background checks on applicants. Partners must use phishing-resistant multi-factor authentication. They may give access only to internal cybersecurity, incident response or penetration-testing teams, must track who uses the model, and may not share, redistribute or sell access.

“Using frontier models can help people to rapidly discover – and fix – critical vulnerabilities,” the programme page says. “But in the wrong hands, the same capabilities can become an equally powerful threat.”

Google said it is also taking part in the U.S. government’s voluntary process for testing models before release. The company did not say which defenders received Carbon this week, or whether any of them are in Africa.

That question matters on the continent. African governments, telecom operators and banks have faced a steady rise in digital crime. Interpol’s Operation Serengeti III, run across 21 countries, led to 1,462 cybercrime arrests. Many public institutions in the region run older software with thin security teams, which is precisely the kind of setting where an automated bug-finder could help most and where an attacker with similar tools could do the most damage.

A pattern across the industry

Carbon’s staged release follows a path other AI companies have started to take. As models grow better at writing and reading code, their makers increasingly hold back the strongest security features or limit who can use them. Chinese startup Moonshot AI said it weakened the hacking skills of its new Kimi model before releasing it openly. Meanwhile the stakes keep showing up in the news: the NSA recently told PHP users to patch two critical flaws or go offline.

Critics argue that “too dangerous to release” also works as marketing. Ars Technica’s headline put it bluntly: Google had announced “its latest frontier AI model that you can’t use.” If Argon is a guide, the wait may be short. That model reached paying subscribers within about two weeks.

When Carbon does open up, it will cost the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens. After the introductory period, Google says, the price doubles to $4 and $20. Google has not said when that period ends.

For now, the people who will judge whether Carbon lives up to its numbers are the defenders Google picked. “We’re eager to see how the initial cohort of cyber defenders use this model,” Kavukcuoglu wrote. Everyone else will have to wait for independent testing, and for Google to decide the guardrails are ready.

Sources

  • Google: Gemini 4 Carbon: our next frontier model (October 28, 2026)
  • Google DeepMind: Fairwind Program
  • The Verge: Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity
  • Ars Technica: Google announces Gemini 4 Carbon, its latest frontier AI model that you can’t use
  • WebProNews: Google Gives Cyber Defenders First Access to Gemini 4 Carbon, Its New Frontier Model
64 Views

By HahuPrime News Desk

The HahuPrime News Desk reports on Ethiopia, Eritrea, the Horn of Africa, East Africa, the Red Sea and the Middle East. Our reports separate confirmed facts from claims and analysis, attribute every claim, and list their sources. Corrections: info@hahuprime.com.