Artificial intelligence is no longer confined to answering questions or writing code. Its ability to plan complex tasks, complete them in multiple steps, verify progress, and correct errors is also growing. Technology is advancing in identifying and fixing software security vulnerabilities. Google’s new model, Gemini 4 Carbon, is an example of this change. However, due to fears of misuse, it is not yet open to everyone.
On 28 October 2026, Google DeepMind announced Carbon, a new member of the Gemini 4 family. Two weeks earlier, Gemini 4 Argon had been released. Google claims Carbon is its most powerful advanced artificial intelligence model to date. Initially, selected cybersecurity experts and defender teams are being given access to use it.
Advanced AI models can plan complex tasks, advance work in multiple steps, and verify results and make corrections as needed. Google engineers used Carbon to carry out complex engineering work lasting several days. This included identifying the root cause of problems in large codebases and creating solutions. This may reduce the need for human guidance at every step. However, human review is essential before verifying and applying results.
Regarding Carbon’s capabilities in cybersecurity, Google said that in their internal tests, it was more than twice as successful at creating secure code compared to Gemini 4 Argon. Google’s Chrome team claims that in creating high-quality security fixes, Carbon performed more than 10 times better than another organisation’s advanced model. These are Google’s published test results, not independently verified comparisons.
In version 5.0 of the Artificial Analysis Intelligence Index, Carbon’s score is 59, the highest on the list. On the same index, Gemini 4 Argon scored 50, Claude Opus 5.5 scored 58, and GPT-6 Astra scored 57. The index evaluates models’ knowledge, complex reasoning, science, coding, and task-completion abilities through various tests. However, although first in overall score, Carbon is not at the top in every type of task.
In the cybersecurity-related SecBench Pro test for identifying software vulnerabilities, Carbon scored 88.2 per cent. In ExploitGym, which tests whether vulnerabilities can be exploited for attacks, its score was 54.8 per cent. And in the CWE Bench for fixing vulnerabilities, its score was 71 per cent. According to information published by Google, Carbon was evaluated in these tests without security measures or guardrails.
By comparison, Argon’s scores were 83.4, 42, and 68 per cent, respectively. Claude Opus 5.5 scored 34.5 per cent on ExploitGym and 67 per cent on vulnerability fixing. GPT-6 Astra scored 85.4 in vulnerability identification, 42.4 in ExploitGym, and 68 per cent in fixing.
Another Google claim is that in a Gray Swan indirect prompt injection test, no attack succeeded against Carbon. However, this is the result of one specific test. It is not a guarantee of complete protection against all types of prompt injection or cyberattacks.
This dual aspect of security is one of the main reasons for limiting Carbon’s use. The ability to identify and fix vulnerabilities can help security teams, but in the wrong hands, it can also benefit attackers. For this reason, Google is limiting Carbon’s use through the Fairwind programme launched in September 2026. The programme was initially launched centring on another cybersecurity model.
Initially, priority is being given to government agencies, organisations operating critical infrastructure in the healthcare, telecommunications, energy, and finance sectors, and trusted cybersecurity partners. Google says there are more than 650 partners worldwide in the Fairwind programme. Later, the situation may change, but for now, Carbon remains behind a carefully guarded door.
Carbon represents both the promise and the caution of advanced AI. It can plan, execute, and correct complex tasks. It can identify and fix vulnerabilities that humans might miss. It can make software safer. But the same capabilities that make it valuable to defenders make it valuable to attackers.
If such tools become more widely available, who will have access? And who will decide? The Fairwind programme suggests that Google is aware of the risks. But the gap between what is possible and what is permitted remains one of the defining tensions of the AI age.
For now, Carbon remains a tool for the few. Whether it becomes a tool for the many depends on how the world navigates the fine line between capability and control.
