Google's Gemini 4 Carbon leads cybersecurity benchmarks

A new standard for defense

Google recently unveiled Gemini 4 Carbon, its latest frontier model designed with a heavy emphasis on cybersecurity and long-horizon capabilities. The model is built to handle complex tasks that require sustained reasoning over many steps. It aims to provide a robust toolset for identifying and fixing software weaknesses.

According to the report from The Hacker News, Carbon has already demonstrated a lead in several key industry benchmarks. It achieved a score of 88.2% on the SEC-bench Pro, which measures the ability to find vulnerabilities. It also performed well on ExploitGym for exploitation tasks and CWE-bench for patching software flaws.

The model is also being positioned as a safer alternative for developers. Google noted that Carbon is twice as likely to generate secure code when compared to its predecessor, Gemini 4 Argon. This suggests a focus on preventing security mistakes during the initial coding phase.

Benchmarking against rivals

The performance of Gemini 4 Carbon has been measured against other leading models in the industry. On the Artificial Analysis Intelligence Index, Carbon earned a score of 59. This puts it ahead of Anthropic's Claude Opus 5.5, which scored 58, and OpenAI's GPT-6.1 Sol, which scored 55.

While Carbon leads in these specific benchmarks, competitors are also making strides in safety and alignment. OpenAI recently released GPT-6.1 Terra, which they claim excels at communication and computer-based tasks. This model is intended to maintain the protections found in the GPT-6 Astra series while staying within user-defined scopes.

Both Google and OpenAI are focusing on making their models more reliable in high-pressure situations. Carbon is designed to seek human input when it encounters ambiguity rather than guessing. Similarly, GPT-6.1 Terra is intended to honestly disclose its own limitations and potential blockers to the user during a task.

Resisting malicious prompts

A major component of Carbon's security profile is its resistance to indirect prompt injection. This is a type of attack where a model is tricked into following malicious instructions hidden within the data it processes. Google claims Carbon successfully resisted every evaluated attempt at this type of attack.

The goal is to ensure that the model remains helpful to the intended user without being hijacked by external inputs. This is particularly important as models are increasingly used to perform agentic tasks, where they can interact with other software and systems autonomously. Security in these workflows is a primary concern for enterprise adoption.

Google is also participating in the U.S. government's voluntary process for pre-release model testing. This indicates a commitment to ensuring that these powerful tools meet safety standards before they are widely deployed in sensitive environments. The company is working to balance high performance with rigorous safety guardrails.

The evolving threat landscape

The release of these models comes at a time when AI's role in cyberattacks is becoming more visible. Anthropic recently shared research showing that an internal model uncovered critical vulnerabilities in major software. These included flaws in Google Chrome, Apple Safari, and the Windows HTTP.sys driver.

The research highlighted that models can now create sophisticated, multi-step exploits in both open-source and proprietary software. Anthropic noted that these skills were previously the domain of specialist teams with deep expertise. Now, models can produce working exploits that anyone can run and verify with very little human involvement.

Because of these risks, Google is limiting the initial rollout of Gemini 4 Carbon to a subset of cyber defenders. Access will eventually expand to other trusted partners as part of the Fairwind Program. This controlled release allows the company to monitor how the model performs in real-world security scenarios.

What comes next for security

The competition between Google, OpenAI, and Anthropic is creating a rapid arms race in AI-driven security. As models become better at finding and patching bugs, the speed of software remediation will likely increase. This could lead to a faster cycle of updates for the software that billions of people use daily.

For organizations, the focus will shift toward integrating these models into existing security pipelines. Defenders will need to learn how to use tools like Carbon to automate the detection of vulnerabilities before they can be exploited by bad actors. The ability to validate and remediate flaws quickly will become a competitive advantage.

As these models continue to evolve, the focus on 'alignment'—ensuring the AI does what the user wants and nothing else—will remain a priority. Both Google and OpenAI are doubling down on making their models more transparent about their limitations. This transparency is crucial for building trust in high-stakes environments like cybersecurity.