Google launches Gemini 4 Carbon for advanced cybersecurity tasks

A new tool for secure coding

Google has officially introduced Gemini 4 Carbon, its newest artificial intelligence model. This specific version is built to excel at complex reasoning, coding, and identifying security weaknesses in software. It arrives just weeks after the company launched the Gemini 4 Argon model.

The new model is designed to help developers write safer code from the start. According to Google, internal benchmarks showed that Gemini 4 Carbon achieved more than twice the success rate of Gemini 4 Argon when generating secure code. This suggests a significant leap in the model's ability to follow security best practices during the programming phase.

The Chrome team at Google also reported impressive results regarding software maintenance. They found that Gemini 4 Carbon delivered more than 10 times better performance than a competing model when generating high-quality code patches. These improvements are intended to help researchers resolve vulnerabilities before attackers can find them.

Benchmark performance results

To prove its capabilities, Google highlighted several third-party and internal tests. The cloud security platform Wiz reported that Gemini 4 Carbon beat other AI systems across various web penetration-testing benchmarks. These tests evaluate how well an AI can autonomously find software flaws.

In the SEC-bench Pro, the model recorded scores of 85.4% and 83.4% for identifying vulnerabilities. It also received a score of 554.8 in ExploitGym, which is a tool used to assess how well AI can validate specific security risks. While the scoring format for ExploitGym requires further confirmation, the results point toward high proficiency.

The model also performed well in Collinear's CWE-bench, where it scored 71%. In this specific comparison, Gemini 4 Carbon reportedly outperformed other models including Gemini 4 Argon, Grok 4.7, and GPT-6 Astra. These benchmarks suggest that Google is specifically targeting cybersecurity professionals who need help with discovery and remediation.

Defending against prompt injection

Beyond finding bugs in code, Gemini 4 Carbon includes improved safeguards against malicious instructions. Prompt injection attacks occur when someone tries to manipulate an AI by embedding hidden commands in user inputs or external content. This is a major concern when AI systems interact with websites or documents.

Google reported that the model reduced unsupported claims by approximately half compared to the Gemini 4 Argon model during internal testing. This indicates the model is becoming more reliable and less likely to provide inaccurate or hallucinated information when faced with complex prompts.

The company also tested the model on Gray Swan's indirect prompt-injection benchmark. Gemini 4 Carbon recorded a 0.0% attack success rate, whereas the previous Gemini 4 Argon model had a 0.7% success rate. While these results are promising, Google notes they reflect specific test conditions and do not guarantee total immunity from all types of attacks.

Rollout and pricing details

Google is not releasing Gemini 4 Carbon to the general public immediately. It is starting with a phased rollout, initially offering access to a select group of cybersecurity experts. This allows the company to monitor how the model performs in specialized professional environments before a wider release.

The company is also coordinating with US government agencies on a voluntary pre-release testing process. This initiative is designed to evaluate the model's capabilities and potential risks before it is deployed more broadly. Additionally, Google plans to provide access to trusted partners through its Fairwind programme.

For those who eventually gain access, Google has set introductory pricing at the same rates as Gemini 4 Argon. To help businesses manage costs, they are also offering a 95% discount on cached input tokens. This discount is intended to lower the price for developers who need to process similar input content repeatedly.