AI hacking skills approach ‘inflection point’

Vlad Ionesco and Ariel Herbert Voss, co-founder Cyber ​​security start RunSybilThey were confused for a moment when… Amnesty International Sybil alerted them to the vulnerability in the client’s systems last November.

Sybil uses a combination of different artificial intelligence Models– as well as some special technical tricks – to scan computer systems for problems that hackers could exploit, such as an unpatched server or a misconfigured database.

In this case, Sybil was tagged problem Through the customer’s deployment of unified GraphQL, a language used to define how data is accessed across the web through application programming interfaces (APIs). This problem meant that the customer was unintentionally revealing confidential information.

What puzzled Ionesco and Herbert Voss was that discovering the problem required a remarkably deep knowledge of many different systems and how these systems interact. RunSybil says it has since found the same issue in other GraphQL deployments — before anyone else reported it. “We looked on the Internet, and it wasn’t there,” says Herbert Voss. “Its discovery was a logical step in terms of the modeling capabilities – a game-changer.”

The situation indicates an increasing danger. As AI models continue to become smarter, their ability to find zero-day bugs and other vulnerabilities also continues to grow. The same intelligence that can be used to discover vulnerabilities can also be used to exploit them.

Dawn songRecent advances in artificial intelligence have produced models that are better at detecting anomalies, says a UC Berkeley computer scientist who specializes in both artificial intelligence and security. Imitative reasoning, which involves breaking down problems into component parts, and agentive artificial intelligence, such as searching the web or installing and running software tools, have enhanced the cybernetic capabilities of models.

“The cybersecurity capabilities of the leading models have increased dramatically in the past few months,” she says. “This is an inflection point.”

Last year, Song co-created a standard called Cybergame To determine how well large language models detect vulnerabilities in large open source software projects. CyberGym includes 1,507 known vulnerabilities found in 188 projects.

In July 2025, Anthropic’s Claude Sonnet 4 was able to find about 20 percent of the vulnerabilities in the standard. By October 2025, a new model, the Claude Sonet 4.5, was able to identify 30 percent. “AI agents can find zero-days, and at a very low cost,” says Song.

Song says this trend shows the need for new countermeasures, including using artificial intelligence to assist cybersecurity experts. “We need to think about how to make AI more helpful on the defense side, and one can explore different approaches,” she says.

One idea is for leading AI companies to share models with security researchers before launch, so they can use the models to find bugs and secure systems before public release.

Another countermeasure, Song says, is to rethink how software is built in the first place. Her lab has shown that it is possible to use artificial intelligence to create code that is more secure than what most programmers use today. “In the long term, we believe this secure-by-design approach will really help defenders,” Song says.

In the near term, the programming skills of AI models may mean hackers will have the upper hand, the RunSybil team says. “AI can create actions on a computer and generate code, and those are two things that hackers do,” says Herbert Voss. “If these capabilities accelerate, it means that offensive security measures will accelerate as well.”

This is an edition of Will Knight Artificial Intelligence Lab Newsletter. Read previous newsletters here.