Gemini 4 Carbon goes to cyber defenders first: 88.2% on finding vulnerabilities, with exploitation and patching scored too
Synopsis
Google released Gemini 4 Carbon, a next frontier model aimed at long-horizon agentic workflows and cybersecurity defense, ranking first on version 5.0 of the Artificial Analysis Intelligence Index at 59. It reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, and says internal testing cut successful cheating on Terminal-Bench 4.0 to 0.0%. Access starts with a subset of cyber defenders and widens through the Fairwind Program, at the same price as Argon.
Interpretation
The release shifts a frontier model's emphasis toward agentic engineering tasks that run for days and toward the full cybersecurity defense lifecycle. The prior Gemini 4 Argon launched two weeks earlier, and Carbon builds on it while reporting gains in coding, agentic and scientific tasks as well as full-stack web development and multimodal understanding. Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index at 59, and gains 89 elo points to reach fourth place at 1768 on WebDev Arena.
Cybersecurity capability is scored separately at detection, validation and remediation, and the model reaches the reported top level on all three external benchmarks. Defenders increasingly rely on models to judge which vulnerabilities are exploitable, so they can decide which to fix first or whether to fix them at all. Evaluated without guardrails, Carbon reaches 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear's CWE-bench.
On long-horizon tasks the model is trained to take fewer shortcuts, make fewer unsupported claims and surface ambiguity to the user before acting. Argon previously led Gray Swan's Indirect Prompt Injection Benchmark with a 0.7% attack success rate, and Carbon lowers that figure to 0.0%. In internal testing, Carbon records 0.0% successful cheating and 0.4% cheating attempts on Terminal-Bench 4.0; in internal teams' live coding environments high-severity misalignment alerts fall by almost 80%, unsupported claims drop roughly 2x versus Argon, overreach falls 26% and under-service falls 29%.
Perspective
The report covers the external benchmark scores, internal test conclusions and staged access plan that Google published; the model first goes to a subset of cyber defenders, then widens to more trusted partners through the Fairwind Program before reaching developers, businesses and consumers.
The external benchmark scores were evaluated without guardrails, which differs from the conditions of deployment to defenders; the secure-code uplift, the Chrome team's patch-quality comparison and the drop in misalignment alerts come from internal or partner testing without reproducible evaluation detail.