Google has unveiled its latest AI model, Gemini 4 Carbon, which the company claims surpasses its competitors in identifying, exploiting, and patching software vulnerabilities. This announcement, made on Wednesday, highlights Google’s commitment to strengthening cybersecurity measures.
Exclusive Access and Fairwind Program
Although not yet widely available, Gemini 4 Carbon is currently accessible to a select group of cybersecurity professionals. Google plans a phased rollout through its Fairwind Program, which provides early access to government entities, critical infrastructure operators, and major tech platforms. Over 650 organizations are already benefiting from this initiative.
The release of Carbon follows the recent launch of Gemini 4 Argon. Some partners within the Fairwind Program have been granted exclusive access to Argon’s cybersecurity capabilities, paving the way for Carbon’s anticipated broader availability.
Benchmark Performance and Capabilities
Gemini 4 Carbon has demonstrated impressive results on various benchmarks. It scored 88.2% on the SEC-bench Pro, which evaluates a model’s ability to identify flaws in real-world software. Moreover, it achieved 54.8% on the ExploitGym exploitation benchmark and 71.0% on the Collinear’s CWE-bench patching benchmark. These scores surpass those of competing models like OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5.
Google emphasizes that these results were obtained without any guardrails, underscoring the model’s potential in real-world applications. The company suggests that understanding exploitability is crucial for prioritizing vulnerability remediation.
Security and Future Outlook
In terms of security, Gemini 4 Carbon excelled in various testing scenarios. It successfully resisted all attacks in Gray Swan’s indirect prompt injection benchmark and showed significant improvements over its predecessor, Argon, in minimizing unsupported claims and high-severity misalignment alerts.
Google is collaborating with the US government in voluntary efforts to test AI models before their public release. The company aims to refine Carbon’s guardrails further based on feedback before making it widely available to developers and businesses.
The introductory pricing for Carbon is set at $2 per million input tokens and $10 per million output tokens. These rates will increase to $4 and $20, respectively, after the initial launch period.
Overall, Google’s Gemini 4 Carbon model represents a significant advancement in cybersecurity, offering robust capabilities for detecting and addressing software vulnerabilities effectively.
