Google's Gemini 4 Carbon AI Shows Advanced Prowess in Cybersecurity Vulnerability Detection and Remediation
Google's latest AI model, Gemini 4 Carbon, demonstrates a significant leap in identifying and fixing security flaws, outperforming its predecessors in internal and external benchmarks.
Google has unveiled its new Gemini 4 Carbon AI model, a sophisticated tool designed to tackle complex software engineering and cybersecurity challenges. Building upon the foundation of its predecessor, Argon, Carbon incorporates enhanced safeguards and advanced capabilities aimed at proactively discovering and rectifying security vulnerabilities within extensive codebases and intricate low-level systems.
Internal testing by Google teams has shown Gemini 4 Carbon dedicating significant time to investigating and resolving bugs, showcasing an impressive ability to handle high-pressure situations and complex analytical tasks. The model's performance in generating secure code has reportedly more than doubled that of Argon in internal benchmarks. Furthermore, Google's Chrome team observed Carbon performing over ten times better at producing high-quality fixes compared to another leading frontier model from a different lab, highlighting its superior efficacy.
In external evaluations, Gemini 4 Carbon has achieved notable scores across several critical cybersecurity benchmarks. It scored 88.2% on SEC-bench Pro for vulnerability discovery in real-world software, 54.8% on ExploitGym for assessing exploitability, and 71.0% on Collinear's CWE-bench for vulnerability remediation. These evaluations were conducted without the model's usual protective guardrails, underscoring its inherent capabilities.
Beyond its detection and fixing prowess, Gemini 4 Carbon emphasizes improved safeguards and reliability. It builds upon Argon's security framework by integrating advanced model training, real-time misuse defenses, robust resistance to prompt injection attacks, and enhanced monitoring of the AI's reasoning processes. Google reports that Carbon made approximately half the number of unsupported claims compared to Argon in internal tests, and it actively explains its limitations and flags discrepancies before acting on ambiguous instructions.
The model also demonstrates a marked reduction in undesirable behaviors. Alerts regarding serious actions conflicting with intended instructions fell by nearly 80% during coding tasks with Google teams. Additionally, there was a 26% reduction in actions exceeding the task's scope and a 29% reduction in what Google terms "under-service." Carbon also successfully resisted all indirect prompt injection attacks tested on the Gray Swan benchmark, a significant improvement over Argon's 0.7% attack success rate.
Google plans a phased rollout for Gemini 4 Carbon. Initial access will be restricted to a select group of cyber defenders, with gradual expansion to additional trusted partners through its Fairwind Program. This approach will allow Google to gather crucial feedback and refine safeguards before a broader release to developers, businesses, and consumers.
Pricing for Gemini 4 Carbon will be set at an introductory rate of $2 per million input tokens and $10 per million output tokens, with a substantial 95% discount on input tokens that are cached. Following the introductory period, standard pricing will be $4 per million input tokens and $20 per million output tokens. This tiered pricing strategy aims to balance accessibility with the advanced capabilities offered by the model.
Google's Gemini 4 Carbon has further solidified its lead in cybersecurity benchmarks, achieving an impressive 88.2% on SEC-bench Pro for vulnerability discovery and 54.8% on ExploitGym for exploitation. The model also demonstrated strong performance in patching, scoring 71.0% on CWE-bench. This latest iteration shows significant gains over its predecessor, Gemini 4 Argon, and surpasses other leading AI models like Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra in key cybersecurity tasks.
The new article provides further details on Gemini 4 Carbon's performance, noting its specific benchmark scores on SEC-bench Pro (88.2%), ExploitGym (54.8%), and Collinear's CWE-bench (71.0%). It also highlights that these results were achieved without guardrails and details its superior performance in secure code generation and patch generation compared to its predecessor, Gemini 4 Argon, and rival models. The article also touches on Carbon's improved safety features, with zero successful attacks in indirect prompt injection benchmarks.