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

Most recent artificial intelligence Models are not not only Pretty good in software engineering-New research shows that they are getting better in finding errors in the program.

At Berkeley, AI researchers in Berkeley’s most recent AI models and agents were the best of 188 large open source codes in 188 large open source codes. To use a car named CybergimAI models have made 17 new mistakes, including 15 or unknown 15 or “zero days”. “Many of these weaknesses are critical,” says the song of the Dawn, the head of the UC Berkeley professor.

Many experts expect AI models to be huge cideration weapons. AI from Startup XBow right now Hackero drove the ranksThe leader board for the Bug Hunter and sits in the upper place right now. The company recently announced $ 75 million in the new finance.

Song has started to change the coding skills of coding skills by improving the coding capabilities of the latest AI models. “It’s a pivotal moment,” he says. “This exceeded our common expectations.”

As the models continue to improve will automate the process of exploiting both discovery and security defects. This can help companies keep the software safe, but can help hackers look at the systems. “We did not try it,” he said. “If we raxed the budget, they were able to make the agents run longer for a longer run.”

UC Berkeley team ordinary border AI models Openai, Google and Antropic, as well as Open source suggestions of Meta, Deepseek and Alibaba, also tested open source suggestions of several agents to find errors Openly, Liverand Mysteriousness.

Researchers used the descriptions of known software from 188 program projects. Afterwards, pads, then analyzed the new codes, tests and concept and analyze the concept exploitation, they fed Percerecurity agents to see that they could identify the same defects for themselves. The team also wanted to hunt for new vulnerabilities from agents.

Through the process, AI tools have created hundreds of proof concept, and these exploits have set two weaknesses and two weaknesses and two patched weaknesses and two of the exploitants. Work, AI, live systems increase the growing evidence that can automate the discovery of zero-day weaknesses, which is potentially dangerous (and valuable) to provide a way to hack.

Nevertheless, it was determined to be an important part of the CyberCurity industry. Security Specialist Sean Heelan recently discovered Openai’s Reasoning Modeling Modeler O3 is a zero-daily defect in Linux kernel with help. Last November, Google declared The project stated that using AI through a program called Zero, using a pre-unknown program.

Like other parts of the software industry, many cyber security companies remain amazed by the AI ​​potential. The new job really shows that AI can regularly find new defects, but also emphasizes the restrictions with technology. AI systems could not find the most shortcomings and stained by especially complex ones.