Google Gemini Carbon Cybersecurity Model, Who Gets It

andImagine a powerful AI that can hunt for software flaws, prove they are real, and write the fix, but almost no one outside Google is allowed to touch it yet. That is the puzzle at the heart of the Google Gemini Carbon cybersecurity model, announced on October 28, 2026. Google reports eye-catching scores for finding, exploiting, and patching security flaws, and it says Carbon tops a major industry index.

Yet access starts only with a small group of vetted cyber defenders, so independent experts cannot yet check those claims. This explainer looks at what Google says Carbon can do, who can use it first, and why its numbers should still be read as claims, not proof.

What the Google Gemini Carbon cybersecurity model is built to do

Google describes Carbon as its strongest model yet, with a focus on cybersecurity and long-horizon engineering work. Long-horizon work means tasks that take many steps over a long period, such as tracing a subtle bug through a large codebase and then fixing it properly. In security terms, Carbon is aimed at the full defence cycle. First, it looks for vulnerabilities in real software. Then it checks whether a flaw can actually be exploited. Then it builds a patch that fixes the root cause rather than papering over the symptom. That focus sets Carbon apart from a general-purpose chatbot, even a very capable one.

How Google says Carbon finds and fixes flaws

Google reported three headline benchmark scores for Carbon, and each one tests a different stage of defence. On SEC bench Pro, which tests finding vulnerabilities in real software, Google reported 88.2 percent. On ExploitGym, which tests whether a reported flaw can be exploited in a controlled test, Google reported 54.8 percent and aslo On CWE bench, which tests turning a validated finding into a working patch, Google reported 71.0 percent. Google also reported that these security scores were run without guardrails, which means the tests measured raw capability rather than performance inside a guarded product. That disclosure is useful, and it also limits what the scores prove about everyday use. Google also reported that Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, just ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57.

• Find real vulnerabilities in real software, reported at 88.2 percent on SEC bench Pro
• Validate risk through controlled exploitation, reported at 54.8 percent on ExploitGym
• Patch flaws at the root cause, reported at 71.0 percent on CWE bench
• Handle long multi-step engineering work across large codebases over days

What the Google Gemini Carbon cybersecurity model costs and who can use it

Access is the most important practical detail, because it shapes everything else in this story. Google is initially restricting Carbon to a subset of cyber defenders through its Fairwind Program, according to reports of the announcement. Access will then widen gradually to trusted partners, but Google has given no public date for developers, businesses, or consumers. If you hoped to try Carbon in a familiar app this week, reports say Google has not named a launch surface or a date for that wider release. For now, Carbon is best understood as a gated preview for defenders, not a product you can buy or try.

Why restricted access matters for trust

Restricted access creates a simple trust problem. The people who could check Google’s scores most credibly cannot yet run the model themselves. That does not mean the scores are wrong. It means they are untested by outsiders, and readers should hold them lightly until more testers get access. Security tools deserve extra care here, because a flaw finder that misses a bug or writes a weak patch can create a false sense of safety. A tool that is genuinely strong can also be misused, which is one reason Google says it is starting with vetted defenders rather than opening the gates at once.

Claim versus confirmed for Gemini Carbon

The table below separates Google’s reported claims from what independent testers have verified so far. Use it whenever a new Carbon headline appears, and ask whether the news adds independent evidence or just repeats a company claim.

Google reports a strong cybersecurity model with impressive benchmark scores and a careful staged rollout. That combination could help defenders find and fix flaws faster, if the scores hold up in real use. For now, the most honest summary is a powerful reported model behind a closed door. The next chapter belongs to the vetted defenders who get access first, and their honest reports will matter more than any launch headline. For related reading, see our cybersecurity automation guide, our AI safety rules, and our Mistral model explainer. External reporting includes the Interesting Engineering report, the Skimwire report, and the Digitbin explainer.