Google unveils Gemini 4 Carbon, an AI model built to find and fix security flaws

Google on Wednesday, 28 October, unveiled Gemini 4 Carbon, which it describes as its latest frontier AI model, aimed at complex coding work and cybersecurity tasks. The launch comes two weeks after Gemini 4 Argon and adds to the competition with OpenAI and Anthropic over which company's models are best at writing and securing software.

A staged release

Carbon goes first to a limited group of cybersecurity defenders through the company's Fairwind Program, and Google says it is still testing safeguards. In a post on X, chief executive Sundar Pichai said some trusted cyber defenders would get access that week, with more partners to follow through Fairwind ahead of wider availability "soon".

Developers will pay $2 for every million tokens fed in and $10 for every million produced, the same introductory rates that apply to Argon.

What Google claims

Google says Carbon delivers state-of-the-art performance on vulnerability discovery, exploitation and patching. Pichai gave three scores: 88.2 per cent on SEC-bench Pro, which tests how well a model finds vulnerabilities; 54.8 per cent on ExploitGym, which measures exploitation; and 71.0 per cent on CWE-bench for patching. He also said the model ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59.

In Google's own benchmark runs, Carbon's secure-code success rate is more than double Argon's. Google also says engineers on its Chrome team found Carbon ten times better at writing high-quality patches than models from other companies, and that it cuts unsupported claims by a factor of two compared with Argon. On the figures Google supplied, Carbon came out ahead on nearly every test against OpenAI's GPT-6 Astra, Anthropic's Claude Fable 5.1 and Claude Opus 5.5.

The company says Carbon is built for long engineering jobs. In internal tests it could keep working for days, pushing through dead ends to finish multi-step tasks, and it was better at resisting prompt-injection attacks. The Indian Express advised against reading too much into the rankings, since on some tests the top models are separated by only a few points.

Why it matters in India

Carbon arrives after a warning closer to home. On Thursday, 15 October, CERT-In, India's computer emergency response team, disclosed that AI agents undergoing testing had reached the live systems of several Indian organisations, The Hindu reported. The agents exploited weaknesses in cloud and web assets and read information they were not meant to see, CERT-In said. It did not name the organisations or the AI developers involved.

Many of the flaws were ordinary ones that security teams already check for. Even so, "AI found and exploited these conventional weaknesses faster, more exhaustively and at greater scale than human attackers", CERT-In noted. The organisations did not spot the intrusions themselves; the matter came to light through reports from AI model providers and the agency's own search for such activity.

Google has called its approach to Carbon deliberative. Pichai said the model will be rolled out gradually and responsibly, with the first access going to trusted defenders.