Google's new frontier model reports leading scores for finding, exploiting, and patching vulnerabilities, plus gains in coding and agentic tasks. Engineers choosing models for agents and security work get new numbers to evaluate.
Key takeaways · AI-distilled
Google is not releasing Gemini 4 Carbon broadly yet: it goes first to a subset of cyber defenders, then to trusted partners via its Fairwind Program, while Google takes part in the US government's voluntary pre-release testing. Developer and consumer access is promised later.
Google reports state-of-the-art results across the defense cycle, evaluated without guardrailsThe checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.Full definition →: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on Collinear's CWE-bench for patching.
Internally, Google says Carbon more than doubles Argon's success rate at generating secure code, and the Chrome team reported it did more than 10x better than another lab's large frontier model at producing high-quality patches.
On alignmentThe work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.Full definition →, Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0% with cheat attempts at 0.4%, high-severity misalignment alerts in internal coding fell almost 80%, and unsupported claims fell about 2x versus Argon.
Carbon launches at the same introductory price as Argon, $2 per million input and $10 per million output tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → with cached input 95% off. Google says internal teams have run it for days on long engineering tasks such as root-causing subtle bugs across large codebases.
Terms in this piece · Glossary
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
