Google Launches Gemini 4 Carbon for Cyber Defense

Mountain View, Oct. 28 — Google on Wednesday introduced Gemini 4 Carbon, its latest frontier model, framing it around cybersecurity defense rather than general chat. The launch lands two weeks after Gemini 4 Argon and as OpenAI, Anthropic, and Google continue to trade coding and agentic releases. Wire services including Reuters summarized the announcement as a bid to lead on vulnerability discovery, exploitation testing, and patching.

Three Benchmarks, Three Jobs

Google organized external results around distinct defender tasks — not a single “security score.” On SEC-bench Pro, which tests autonomous bug finding in real software, Google reports 88.2%, up from Argon’s 83.4%. On ExploitGym, which checks whether a flaw can be turned into a working exploit, Carbon hits 54.8% versus 42.0% on Argon — the largest step-up in the set. On Collinear’s CWE-bench, which measures root-cause patches, Google cites 71.0%, against 68.0% for Argon.

Those numbers are useful only with the caveat Google attached: the public cybersecurity evaluations ran without guardrails. Customer-facing deployments will ship with misuse filters, prompt-injection defenses, and policy layers that can cap the same tasks. VentureBeat and other outlets flagged the gap between leaderboard runs and production behavior as the central interpretive issue — finding a suspicious line of code is not the same as proving exploitability, and proving an exploit is not the same as landing a merge-ready fix.

Internally, Google claims more than 2× Argon’s success rate on secure-code generation and says the Chrome team saw 10× better high-quality patches than an unnamed rival lab model. Wiz, Google’s cloud-security unit, pointed to web penetration-testing benchmarks tied to its Scan for Good program. Those anecdotes are directional; they are not independently reproducible from the blog post alone.

Gated Rollout Is the Product

Carbon ranks #1 on version 5.0 of the Artificial Analysis Intelligence Index at 59, Google said, and adds +89 Elo on WebDev Arena to reach #4 overall — evidence the model is not cyber-only. Still, Google is not opening API access broadly yet. It is engaging the U.S. government’s voluntary pre-release testing process, then phasing access through the Fairwind Program to cyber defenders and trusted partners while guardrails iterate.

That makes platform control the binding constraint: the same organization that publishes scores also decides who may run unconstrained variants. Pricing matches Argon’s introductory $2 per million input tokens and $10 per million output tokens, with cached input at 95% off; after the intro window, Google lists $4 / $20. For enterprise security teams, the invoice matters less than whether Carbon arrives inside Google Cloud, Chrome, and Wiz pipelines before rivals embed comparable agents in their stacks.

Google also emphasized alignment metrics carried forward from Argon: 0.0% successful cheating on Terminal-Bench 4.0, roughly 2× fewer unsupported claims in internal tests, and 0.0% success on Gray Swan’s indirect prompt-injection benchmark. Those figures describe Google’s internal harness, not every customer environment.

What Practitioners Should Watch

For defenders, Carbon’s pitch is cadence: autonomous triage across find–validate–patch maps directly onto workflows that still bottleneck on human reviewers — the same structural pressure described in recent AI-and-incident coverage, where patch speed rather than model novelty sets the risk.

For the industry, dual-use is explicit. ExploitGym gains are offensive capability dressed in defensive packaging. Google’s answer is tiered access and government review — a STATE-adjacent checkpoint on a PLATFORM-owned model.

What would change the verdict: published third-party replication under production guardrails; Fairwind cohort case studies with time-to-patch data; or a general-developer release that shows whether Carbon’s SEC-bench lead survives outside Google’s evaluation harness.

Until then, Carbon is less a consumer product than a statement of intent: the frontier model race now includes autonomous vulnerability work — shipped first to the defenders Google selects.