Google has unveiled Gemini 4 Carbon, a new frontier model that it says leads benchmark tests for finding software vulnerabilities, validating exploits and generating patches. The company announced the model on Oct. 28, describing it as its strongest system for cybersecurity and long-horizon engineering work.
According to Google’s announcement, Carbon scored 88.2% on SEC-bench Pro, which measures the discovery of flaws in real-world software. It scored 54.8% on ExploitGym, an evaluation of whether models can determine if vulnerabilities are exploitable, and 71.0% on Collinear’s CWE-bench for producing code fixes. Google says the external cybersecurity evaluations were conducted without guardrails.
The results span the principal stages of vulnerability response: identifying a weakness, establishing its practical risk and addressing its root cause. That combination could help security teams triage large codebases and accelerate remediation, but the ExploitGym result also highlights the dual-use risk of giving frontier models stronger exploit-development capabilities. Google has not tied the announcement to a specific CVE, public proof of concept, exploitation campaign or threat actor.
Google says Carbon is built on the safety and security controls used for its predecessor, Gemini 4 Argon, including model training, real-time misuse defenses, prompt-injection resistance and chain-of-thought monitoring. The company reports that Carbon resisted all evaluated attacks in its internal indirect-prompt-injection test and reduced high-severity misalignment alerts in live coding environments, claims that remain based on Google’s own testing rather than an independently published security audit.
Access is initially limited to a subset of cyber defenders, with wider distribution planned through Google’s Fairwind Program and additional trusted partners. Google says it is participating in the U.S. government’s voluntary pre-release model-testing process and will continue refining safeguards before making Carbon available to developers, businesses and consumers. Until broader testing is published, defenders should treat the benchmark figures as vendor-reported indicators, validate model-generated findings and patches in isolated environments, and require human review before any exploit or remediation reaches production.
