Google has officially begun the targeted rollout of its Gemini 4 Carbon model. Starting this week, the cutting-edge large model will be made available to a group of cybersecurity defenders through the Fairwind program. Its AI capabilities will help security teams automatically discover, verify, and patch software vulnerabilities. Google is also enforcing strict access controls to prevent malicious hackers from exploiting the model’s powerful capabilities.

Gemini 4 Carbon is Google’s next-generation frontier model, following Gemini 4 Argon. It has delivered impressive results in code analysis, autonomous agent tasks, and specialized cybersecurity vulnerability discovery tests. It scored 88.2% on SEC-bench Pro, a vulnerability-related benchmark, and 71.0% on CWE-bench, a benchmark for vulnerability remediation—improving on its predecessor’s ability to identify system flaws and generate effective patches. Security defenders can use Carbon with the CodeMender tool to scan target systems in batches, quickly locate critical vulnerabilities, and generate code to fix them. This can greatly reduce the time and labor required to find and remediate vulnerabilities, making it especially suitable for cybersecurity work at high-risk organizations such as critical infrastructure operators, government agencies, telecommunications companies, and healthcare providers.

Google’s choice of the Fairwind program as the first deployment channel for Carbon was no accident. The program is specifically designed for national cybersecurity authorities, critical infrastructure operators, and trusted security partners that have undergone due diligence. Its core goal is to put cutting-edge AI models in the hands of cyber defenders first, helping them strengthen their systems before hackers use similar AI tools to launch attacks. Google has imposed strict usage restrictions: partners enrolled in the program may use the models only for activities such as defensive security research, malware analysis, and authorized threat simulations, preventing their misuse to develop attack programs or malicious exploit code.

This phased rollout also reflects Google’s careful balancing of safety considerations in releasing cutting-edge AI capabilities. Carbon has powerful code-generation and vulnerability-analysis capabilities, which can be used for both good and harm: they can help protect networks, but in the hands of attackers, they could also be used to find vulnerabilities across systems on a massive scale and carry out larger cyberattacks. That is why Google will not make Carbon immediately available to all developers or ordinary users. Instead, it is first being rolled out to a small group of cyber defenders, while Google gathers feedback and improves the model’s safety guardrails. Access will be expanded gradually only after the risks are deemed manageable.

The broader industry landscape shows that cyber offense and defense have entered a new, AI-driven era. Hackers have already begun using large language models to help write malware and find website vulnerabilities. Meanwhile, tech giants such as Google and Microsoft are trying to put their most powerful AI tools in the hands of defenders, pursuing a “defense first” strategy. Those who can use AI to find vulnerabilities at scale sooner will gain the upper hand in cyber conflicts. The targeted rollout of Gemini 4 Carbon means Google is extending the competition over large language models onto the cybersecurity battlefield.

Key areas the market will be watching include the real-world impact of Carbon on vulnerability remediation in enterprise environments, the pace at which the Fairwind program expands its partner base, and when Google will open up API access to Carbon for ordinary business customers. Meanwhile, regulators around the world will continue to monitor how highly capable large language models are governed. The dual-use risks of AI models will be an unavoidable issue in future AI regulation.