Google Unveils Gemini 4 Carbon, a Frontier Model Aimed at Cyber Defence
Seven Lights
Google announced Wednesday that Gemini 4 Carbon, its newest frontier AI model, leads rival systems at detecting, exploiting and patching software vulnerabilities, and said it would restrict access to vetted defenders rather than release the model openly. The company opened access this week to a select group of cyber defenders and will expand it in stages to a wider pool of approved partners through its Fairwind Program, which offers governments, critical infrastructure operators and major technology platforms a path to advanced models.
Google’s claims rest on benchmarks. On SEC-bench Pro, which tests a model’s ability to uncover flaws in real-world software, Carbon led competing systems, the company said. On its internal measures, Carbon was more than twice as successful as its predecessor Gemini 4 Argon at writing secure code. Google’s Chrome team found that when generating high-quality patches, Carbon outperformed a large frontier model from a rival laboratory by more than ten times. Wiz, the Google-owned cloud security firm, observed Carbon take the top spot on web penetration-testing benchmarks.
Google framed the exploitation of vulnerabilities as a defensive step, one that helps security teams decide what to fix first, or whether to fix at all. “To prioritize which vulnerabilities to remediate first or remediate at all, cyber defenders increasingly rely on models to check which vulnerabilities are exploitable,” the company said.
The gating is the story’s sharper edge, Ars Technica noted in headlining that the model “you can’t use.” A model that can find and exploit flaws at scale is also a model whose open release could arm attackers, and Google’s answer — a managed program with approved partners — follows the pattern set when Argon arrived two weeks ago, some Fairwind partners holding exclusive access to its defensive capabilities in the meantime. Mashable and CNET covered the announcement as part of Google’s broader frontier push, while Engadget reported that Carbon promises a significant leap in AI coding and cybersecurity. The Register called the model a diamond in the rough.
Whether the restricted-access model holds commercially depends on whether the approved pool moves fast enough for the organisations that need it. Google said the expansion through Fairwind would proceed in stages.