AI Agents Are Getting Better at Writing Code—and Hacking It as Well

Last Artificial intellect Models are not just Software is significantly better in engineeringNew research shows that they are always selling bugs in software.

AI researchers from UC Berkeley have examined that the latest AI models and agents can find vulnerability in 188 large open source codebase. Using a New benchmark Has been called CyberjimAI models have identified 17 new bugs with 15 previously unknown, or “zero-day”. “Many of these weaknesses are critical,” said Dawn Song, Professor of UC Berkeley, who led this work.

Many experts hope that the AI ​​models will turn into cyberquacy weapons. Currently an AI tool from Startup XBO Hacerone Rank has made upLeaderboard for bug hunting and currently sitting in the top place. The company recently announced $ 75 million of the new fund.

The song says that the coding skills of the latest AI models combined with the improvement of logic skills have begun to change the cyberquacy landscape. “This is an important moment,” he says. “It has actually exceeded our general expectations.”

Models continue to improve as they do Protection errors will automatically automatically discover and absorbThe It can help companies protect their software but can help hackers break the system. The song says, “We didn’t even try.” “If we are spreading the budget, allowed agents to run for more time, they can do better.”

UC Berkeley Team OpenAEEE, Google and Ethnic Border AI models have tested the conventional AI models, as well as open source offers from meta, dipsek and Alibaba to find the bugs, including the bugs. Open -handed, CybcheAnd EnigmaThe

Researchers used details of the software weaknesses known from 188 software projects. They then feed the details of the cybercuity agents powered by Frontier AI models so that they analyze new codebuses, conducting tests and creating proof-off-concepts absorbing the same defects for themselves. The team asked agents themselves to look for new vulnerability to codebase.

Through the process, AI equipment created hundreds of proof-off-concept absorptions, and in which the researchers identified 15 previously unseen weaknesses and two weaknesses that were previously published and patched. The work has added to the growing evidence that AI can automatically automatically invention of zero-day weaknesses, which are possible dangerous (and valuable) because they can provide a way to hack live systems.

AI still seems to be an important part of the industry in the industry. Security expert Recently discovered A zero-day defect in the Linux kernel, widely used with the support of the Openai logic and 3. Last November, Google Declaration It was discovered by a program called AI through a program called Project Zero.

Like the other parts of the software industry, many cybersic security companies are fascinated by the possibilities of AI. The new job actually shows that AI can regularly find new defects, but it also highlights the rest of the restrictions with technology. AI systems were unable to find most errors and were stumped by especially complications.