Google Says Gemini 4 Carbon Is a Cybersecurity Leap—But You Can’t Use It Yet
Google Says Gemini 4 Carbon Is a Cybersecurity Leap—But You Can’t Use It Yet
Google’s latest frontier-model push began only two weeks after the release of Gemini 4 Argon. With Gemini 4 Carbon, the company is now claiming a sharper focus on coding agents, scientific work and, above all, cybersecurity.
Carbon is being marketed as a “significant leap” in finding, validating and fixing vulnerabilities. Google says it is twice as likely as Argon to produce code without vulnerabilities, while Chrome engineers say it generates high-quality patches at 10 times the rate of non-Google models. The company also says Carbon withstood every prompt-injection attempt in its Grey Swan evaluation.
But the rollout is as notable as the claims. Google is taking what it calls a “deliberative approach,” testing through the US voluntary framework before releasing Carbon via its Fairwinds cyber-defense program. There is no firm timetable for broad availability; early access is expected to favor AI Ultra subscribers and Gemini API developers.
Sundar Pichai cast the rapid follow-up to Argon as evidence of Google’s momentum, saying Carbon “shows the pace at which we’re pushing the frontier.” Demis Hassabis struck a similar note, calling the model state of the art across major benchmarks and saying it combines frontier capability with “deep safety & security research.”
The more cautious reading is that Google’s numbers do not settle the race. Carbon reportedly edges rivals including GPT-6 Astra and Claude models in most of Google’s tests, but the margins are only a few points and benchmark results cannot fully capture real-world performance. Even so, the launch earned a concise endorsement from Elon Musk: “Impressive.”
For now, Carbon’s central tension is plain: Google is presenting it as a security breakthrough while asking the public to wait for proof outside its own controlled evaluations.
Write a comment