AI-generated conceptual illustration of cybersecurity and software repair
Finding a software vulnerability, checking whether it matters and fixing its underlying cause are separate tasks. Google is positioning Gemini 4 Carbon across that sequence—and initially limiting who can use it.
Announced on October 28, Carbon is Google’s newest frontier model. The company says it improves coding, cybersecurity and long-running agent workflows compared with Gemini 4 Argon, introduced two weeks earlier.
Capability and access move on different schedules
Google reports scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, covering vulnerability discovery, exploitation and patching respectively. The announcement notes that these external benchmark evaluations were conducted without guardrails.
Those results describe performance under particular evaluation conditions. They do not establish the same success rates in customers’ software environments.
Starting this week, Google plans to provide Carbon to a subset of cyber defenders, gradually extending access to trusted partners through its Fairwind Program. Wider access for developers, businesses and consumers remains forthcoming.
The practical test is a verified repair
For AIYXL, the useful deployment question is whether a model can carry evidence through the entire repair process.
A convincing vulnerability report should identify the affected code and explain the conditions required to trigger the flaw. A proposed patch should address that cause, preserve intended behaviour and survive relevant tests. Producing more findings helps only if teams can determine which ones deserve action.
The distinction becomes especially important in sustained autonomous work. A model may make progress across many steps while introducing an incorrect assumption early in the process. Reviewable intermediate results give engineers a way to detect that drift.
Carbon’s release therefore creates two things to watch: how its reported capabilities translate into verified security work, and how Google expands access after the initial restricted deployment.