Google has introduced Gemini 4 Carbon, a new AI model designed for in-depth cybersecurity work—spotting software flaws, testing whether they’re exploitable, and generating effective fixes. Unveiled on October 28, 2026, Carbon builds atop Gemini 4 Argon, aiming to dramatically cut down the time it takes to tackle complex security and coding challenges. Access begins with selected cyber defenders rather than a full public rollout.
Benchmark Gains & Security Testing
On internal evaluations, Gemini 4 Carbon achieved strong results across vulnerability spotting, exploitation, and patching. It scored 88.2% on SEC-bench Pro (vulnerability discovery), 54.8% on ExploitGym (exploitation), and 71.0% on CWE-bench (patching). These cover three key stages in cybersecurity work: detecting flaws, assessing exploitability, and delivering fixes for the underlying issue.
Beyond security tests, Carbon topped the Artificial Analysis Intelligence Index v5.0 with a score of 59, a metric evaluating performance across coding, science, knowledge work, and general knowledge. It reportedly outperformed Argon in secure code generation by more than double and—when it came to producing high-quality patches—eclipsed an unnamed competitor by over tenfold, though details of that rival were not shared.
Safety Measures & Access Strategy
To protect against misuse, Carbon is built with a multi-layered defense strategy. It includes checks against misuse, safeguards against prompt injection, and internal monitoring of reasoning and actions. On the Gray Swan Indirect Prompt Injection benchmark, Carbon had an attack success rate of 0.0%, compared to Argon’s 0.7%. That said, Google notes this result applies only to the evaluated attacks—and real-world risks remain.
Other metrics show progress: high-severity misalignment alerts on internal coding tasks dropped by nearly 80%, and unsupported claims were cut roughly in half versus Argon. In addition, Carbon achieved a zero success rate for cheat attacks in the reported Terminal-Bench 4.0 evaluation.
Carbon is initially being made available to a select group of cyber defenders, followed by broader access via Google’s Fairwind Program and through voluntary U.S. government testing. Launch pricing is set at $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95%. Later, standard rates will rise to $4 and $20 respectively. Google has not yet named a firm date for general release.
Why This Matters
Gemini 4 Carbon could mark a turning point in AI-driven security. Detecting vulnerabilities is only step one; the harder work lies in assessing impact and producing reliable solutions at scale. By demonstrating strong benchmark performance across discovery, exploitation, and patching—and investing in safety layers—Google is leaning into the enterprise need for trustworthy, end-to-end security tooling.
Enterprises and defenders should keep watch on how Carbon performs in real environments: its benchmark success will need to translate into consistent, safe results in diverse codebases and under adversarial conditions. Upcoming participation in government testing and the Fairwind Program may offer clearer evidence of readiness. The cost structure suggests that carbon usage may be costly at scale unless optimization or caching is in place. Still, Carbon raises the bar on what AI in cybersecurity is trying to achieve—and could force competitors to chase similar holistic models.
