Google calls Carbon its strongest model for cybersecurity and long-horizon work and says it ranks first on Artificial Analysis Intelligence Index v5.0 (59). Security-relevant claims from the launch post:

• Defense lifecycle: 88.2% on SEC-bench Pro (finding), 54.8% on ExploitGym (exploiting), 71.0% on Collinear's CWE-bench (patching). More than 2x the secure-code success rate of Gemini 4 Argon on internal tests; the Chrome team reports Carbon did more than 10x better than "another lab's large frontier model" at high-quality patches.
• Prompt injection: 0.0% attack success on Gray Swan's indirect prompt injection benchmark (Argon was 0.7%).
• Rollout: starting this week for a subset of cyber defenders, then more trusted partners through the Fairwind Program, before general availability. Google says it is taking part in the U.S. government's voluntary pre-release testing. Launch price $2 / $10 per million input / output tokens.

Take (opinion): a model that scores this well on exploitation is useful to attackers too, which is why access is gated. A 0% result on one prompt-injection benchmark is not a guarantee in your own agent; keep tool permissions and egress controls in place.