The News, Explained

Correction (November 10, 2026): We corrected the original announcement date to October 28 and anchored its relative access timeline to that announcement week, following the official source’s current address and date. OYOPICK first published this article on November 10. This correction concerns the source date and the timing of the original plan. Source

Google announced Gemini 4 Carbon on October 28, 2026. At that time, it said selected cybersecurity defenders would receive early access during the week of that announcement, with access set to expand through trusted partners in the Fairwind Program. That announcement described availability through products for developers, businesses and the public as a future plan, without a specific general-release date. Source

Carbon focuses on coding and sustaining multistep work over longer periods. Agentic work means an AI uses tools to investigate a problem and decides what to do next from the results. Google says its internal teams observed Carbon persisting for days on complex tasks involving large codebases, investigating bugs and producing patches. This is an observation reported by the company from its tests. Source

Its security capabilities can be divided into finding vulnerabilities, checking their practical risk and producing patches that address their causes. For example, discovering a flaw in a program and checking whether corrected code works are different stages. Google reports 88.2% on SEC-bench Pro, a vulnerability-discovery evaluation. It specifies that the external cybersecurity results were measured without guardrails, the safety controls that constrain how a model handles requests and acts. Source

Google describes combining training, real-time abuse prevention, defenses against misleading external instructions and monitoring of the work process. It reports a 0.0% attack success rate in an indirect prompt-injection evaluation. That measures the evaluated attacks in which malicious instructions are placed in material the model reads to redirect its work. Source

The Fairwind page provides background on access procedures. The program prioritizes defenders of critical services such as government, healthcare and telecommunications, and checks applicants’ security history and ethical operations. Applications go through a Cloud account team. Access is restricted to approved partners, who cannot resell or redistribute it. The current program page explicitly names Gemini 4 Argon as its model; distinguish that existing guidance from the October 28 announcement’s planned expansion to Carbon. Source Source

OYOPICK’s Take

We welcome the direction of AI extending the capabilities of people who protect important services by investigating problems longer and helping fix flaws. For an adoption decision, we think it is more useful to ask which stage of the work benefits than to rely on one score. With vulnerability discovery and patch generation evaluated separately, finding a flaw and reviewing its proposed fix also deserve separate attention.

Disclosing both performance tests without guardrails and limited access for specialists helps readers distinguish capability from deployment conditions. A 0.0% result describes what was observed in that evaluation; the evidence does not justify extending it to every external document or future attack. As access expands, we hope Google specifies the supported tasks and accompanying restrictions, making the benefits for defenders easier to explain to a wider public.