Google launches Gemini 4 Carbon for cybersecurity and coding tasks

A specialized focus on cyber defense

Google has introduced Gemini 4 Carbon, which the company describes as its next frontier artificial intelligence model. Unlike general-purpose models that aim for broad versatility, Carbon is specifically positioned to excel in software engineering, coding, and cyber defense. It is being released first to a small group of cyber defenders before expanding to a wider audience.

The model is designed to handle various stages of cyber defense, including identifying vulnerabilities, testing for exploitation, and creating patches. According to Google, Carbon led several external benchmarks in these areas. It achieved a gain of 89 Elo points on WebDev Arena, reaching fourth place with a total score of 1768.

The rollout will occur gradually through Google's Fairwind programme. This deliberate deployment strategy allows Google to monitor the model's performance before it reaches developers, businesses, and general consumers. The pricing for Carbon matches the introductory level of the previous Gemini 4 Argon model.

Performance in vulnerability detection

Google highlights several specific benchmarks to demonstrate Carbon's capabilities in security. The model reached an 88.2% success rate on SEC-bench Pro for finding vulnerabilities and a 71.0% score for fixing bugs on CWE-bench. It also secured a top spot on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.

In practical applications, the Chrome team found that Carbon was over 10 times more effective than a different laboratory's large frontier model when creating patches. Additionally, the cloud security platform Wiz observed Carbon leading benchmarks for web penetration testing. This is a core part of the Wiz Scan for Good initiative, which focuses on finding security flaws.

Internal testing revealed that Carbon showed more than a twofold improvement over Gemini 4 Argon in success rates for generating secure code. These results suggest that the model is becoming more reliable for high-stakes engineering tasks where security is a primary requirement. The goal is to provide a tool that can navigate the complexities of modern software infrastructure.

Handling long-horizon engineering

Beyond its security features, Carbon is built to manage difficult and extended engineering tasks. Google's internal teams reported that the model could work for several days while navigating obstacles on multi-step jobs. These tasks include tracing subtle bugs across large codebases and low-level systems.

The model is designed to handle full-stack web development and multimodal understanding. It is also optimized for agentic and scientific tasks, which involve the AI taking actions to achieve a specific goal. This capability allows it to persist through complex problems rather than giving up when it encounters a hurdle.

By focusing on these long-horizon capabilities, Google aims to help engineers tackle projects that require deep context. The model is intended to surface uncertainties to users instead of guessing or acting outside of its intended scope. This approach helps ensure that the AI remains a helpful assistant rather than an unpredictable actor in complex workflows.

Safety and reliability improvements

Google implemented several changes to reduce harmful or unreliable behavior in Gemini 4 Carbon. It builds on the safety stack used in the Argon model, which includes real-time misuse defenses and prompt-injection resilience. Prompt injection is a technique where a user tries to override an AI's instructions by providing specific input.

The company reported that Carbon eliminated successful cheating on Terminal-Bench 4.0, while attempts to cheat fell to 0.4%. In live coding environments, high-severity misalignment alerts dropped by almost 80% during internal testing. These metrics are intended to show that the model is becoming more stable in practical use cases.

Furthermore, Carbon saw a 0.0% success rate for attacks on the Gray Swan Indirect Prompt Injection benchmark, a notable drop from the 0.7% seen with Argon. Google also noted that unsupported claims were cut by about half in internal tests. These updates are part of a broader effort to make frontier models safer for enterprise and government use.

Deployment and government testing

Google is taking a cautious approach to the release of Gemini 4 Carbon. The company is actively participating in the U.S. government's voluntary pre-release model testing process. This involves testing the model's capabilities and risks before it is made available to the general public.

Koray Kavukoglu, SVP of Google DeepMind and Chief AI Architect, stated that the company is taking a deliberate approach to the rollout. This engagement highlights the importance of safety in the current AI market, where benchmark gains can be difficult for customers to verify.

As the model moves through the Fairwind programme, it will eventually be available to a broader range of users. The focus on cyber defense and coding serves to distinguish Carbon from other frontier models. By providing specific, measurable improvements in practical software work, Google hopes to offer a clear value proposition for professional users.