Gemini 4 Carbon arrives as a defensive model while security chiefs put AI at the top of their 2027 agenda

What Google released

Google describes Gemini 4 Carbon as its strongest model to date, built for long-horizon engineering work and for cybersecurity defense. That framing matters because it positions the model as a tool for defenders first, rather than as a general release for everyone at once. m:…3863

Access is being phased through the Fairwind Program, and Google reports that a subset of cyber defenders gets access starting in the week of October 28, 2026, with additional trusted partners to follow through the same program. The phased rollout means most developers will watch the first deployment before they can test the model themselves. m:…5373

Google prices the model at 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens priced at 95 percent off the input token price. For teams running long defensive workflows, where the same context gets reused across many checks, the cached-input price is the number that shapes the real bill, and it is the part of the pricing worth modelling before committing to a workflow built around this model. m:…5373

The benchmark case Google makes

Google reports scores of 88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym and 71.0 percent on CWE-bench, and it reports that the model ranks first on the Artificial Analysis Intelligence Index version 5.0. Those are vendor-reported results, so they should be read as the case Google is making for the model, not as independent verification, and the scores cover vulnerability detection, exploitation and patching work rather than general capability. m:…3863

Google also reports a 0.0 percent success rate on the Gray Swan Indirect Prompt Injection Benchmark and a jump of 89 elo points on WebDev Arena. The injection result is aimed at a specific worry for defensive deployments, namely whether an outside input can steer the model, while the WebDev result speaks to the long-horizon engineering work the model is designed to sustain. m:…5373 m:…3863

Google notes that it is engaging in the United States government voluntary process for pre-release model testing. That detail matters for buyers in regulated settings, because a model sold for defensive work will be judged on how it was tested before release as much as on its benchmark scores. m:…3863

What security chiefs say they will buy

A survey of 100 chief information security officers reported by Andreessen Horowitz finds that 87 percent expect cybersecurity spending to increase in 2027, and that all seven of the most-selected security initiatives for 2027 involve AI. Nearly half expect security headcount to increase in 2027 as well, which suggests that AI tooling is being added to teams rather than replacing them. m:…9797

The same survey finds that 63 percent of enterprises report that security concerns are slowing AI adoption, while 92 percent report at least limited production use of employee generative AI tools. Only 18 percent report production use of AI security operations tools. That gap between broad employee use and narrow production security use is the opening a defensive model is aimed at, and it is also the reason buyers will ask for deployment evidence rather than benchmark scores alone. m:…9797

The survey also reports that 52 percent of chief information security officers expect offensive security testing to run continuously against production assets within a year, and 60 percent expect to be comfortable with autonomous agent actions on low-risk tasks within 12 months. Bain estimates cited in the same article put the compression of a typical attack sequence from four weeks to about eighteen hours. Taken together, those numbers describe a market where testing becomes continuous and agent actions become routine, which raises the stakes for getting the defensive tooling right. m:…9797

Why this pairing matters

This suggests a two-sided race rather than a single product launch. On one side, a frontier model is being sold on its ability to find and patch vulnerabilities. On the other, security leaders say they plan to spend more, test continuously and let agents act on low-risk tasks, while also saying that security worries are slowing adoption. A model that arrives through a phased defender program fits that caution, because it lets a small group of defenders test it before wider release. m:…5373

The practical question for an executive is not whether the benchmark scores are high. It is whether the model holds up inside a production security workflow, where inputs are adversarial, context is reused and mistakes have consequences. The phased rollout, the cached-input pricing and the survey results all point to the same test: run it on a narrow defensive job, measure it against the existing stack, and expand only if the results survive contact with real traffic. m:…9797