Google has introduced an artificial intelligence system that could detect security vulnerabilities that are particularly difficult to find.
Software can contain an error that no one notices for years. Yet this error can sometimes allow a hacker to steal data or take control of a computer. Before they are discovered by malicious individuals, security specialists try to find these types of errors so they can be fixed.
This is precisely one of the areas where Google says it has made progress with Gemini 4 Carbon, announced on October 28. This new artificial intelligence model is not only used for chatting or writing text. It is also designed to carry out computer tasks that require many steps.
In practical terms, such an AI can examine a program’s code, meaning the instructions that make it work. It looks for errors that could create an opening for a hacker. It can then test whether the problem actually poses a risk and suggest a fix. This work usually takes software protection teams a great deal of time.
To measure Carbon’s progress, Google cites several series of tests (benchmarks). In one of them, focused on finding vulnerabilities in real software, the model achieves a score of 88.2%. In another, focused on fixing the problems identified, it reaches 71.0%.
The company also reports results obtained by its own teams. According to its tests, Carbon succeeds more than twice as often as Gemini 4 Argon in producing flawless code. Argon is another model that Google had presented just two weeks earlier.
These announcements nevertheless call for caution. A good test score does not mean that an entire piece of software will automatically be secure. Google’s internal results are not an independent verification either.
There is another difficulty: understanding how to exploit a vulnerability could also help malicious individuals. That is why Google is not immediately making Carbon available to everyone. The company says it has strengthened its safeguards and plans a gradual rollout rather than a general launch as soon as it is announced.
At first, only certain professionals responsible for securing computer systems should be able to use it. Their feedback will allow Google to adjust the safeguards and continue its evaluations. No precise date has yet been given for making it available to the general public.