AI cyberattacks are changing how quickly software weaknesses get found, and defenders are feeling the squeeze. Researchers say agentic AI tools can comb through enormous amounts of code in minutes, which gives attackers a head start before organizations can fix a flaw.
The warning follows a run of disruptions at Microsoft, US air traffic control, JPMorgan Chase and even the US National Security Agency. It also follows reports of AI-assisted bug discovery in Java and PHP software, and of hackers using AI tools against supermarket fridges and freezers in Europe.
Why the speed matters
Steven Murdoch, a computer security researcher at University College London, says finding vulnerabilities without AI is essentially a thing of the past. Researchers who skip the technology risk falling behind.
Before AI agents, a hacker might spend hours inspecting a website and probing software for a way into a network. Agents now run automated tools that can search millions of lines of code. Google has reported that attackers have already used agentic AI to study software updates, work out which security problem each one fixes, and then build ways to exploit it.
Daniel Gardham, a cybersecurity researcher at the University of Surrey, says the real change is scale. An attacker can start an agent, step away for 20 minutes and return to dozens of findings. AI tends to find the same kinds of weaknesses human hackers do. It simply lets people with limited technical skill reach them.
Murdoch adds that the tools never tire, so a team can run a hundred of them side by side. They are especially good at familiar categories of flaw. The most accomplished human researchers still do better at spotting entirely new kinds of weakness.
People remain the weak point in AI cyberattacks
Gardham stresses that most breaches still come down to human failure. The European freezer incidents are a good example: many devices were left with default passwords such as “admin” or “12345”.
AI also makes phishing harder to spot. Attackers can feed it publicly available information to personalize fraudulent messages, which makes them look authentic and more likely to succeed. Clicking a link from an unfamiliar sender remains one of the simplest ways in.
The Java and PHP finds
AI helped uncover security flaws in Java and PHP, two languages used to build cross-platform software and websites. In the Java case, the AI identified conditions that could let an attacker corrupt computer memory.
Murdoch thinks that issue has been overblown. It is a genuine problem, he says, but nowhere near the top ten this year, and probably not even this week.
Defenders use the same tools
The picture isn’t one-sided, though. Defense against AI cyberattacks increasingly relies on AI as well.
Security researchers are using AI to hunt for weaknesses too. Murdoch says this has become increasingly important to people who find vulnerabilities, because they can now examine far more software and work out how to patch it. Teams are also building AI agents that can detect and repair flaws on their own.
The usual process is that a researcher who finds a weakness tells the developer privately, so engineers can prepare a fix before the details go public. AI has compressed what comes after. Murdoch says the gap between public disclosure and attempted exploitation has shrunk from several weeks to just a few days.
Rushing a fix has costs. A patch can interfere with other programs and cause the kind of outage that organizations are trying to avoid. Murdoch says the job of the defenders trying to fix things has become much harder.
Were the outages connected?
Probably not. Murdoch considers it unlikely that the recent outages were a coordinated large-scale attack. He believes the most prominent disruptions were more likely technical malfunctions than anything sinister, even though they have renewed concerns about how resilient critical digital infrastructure really is.
What researchers want to learn next
Researchers now want to know how thoroughly AI can check software for security flaws, and whether it could introduce new weaknesses of its own. For organizations, the practical message is shorter patch cycles. If exploitation can follow disclosure within days, a fix that waits weeks is exposed for most of that time.
Basic hygiene still counts against AI cyberattacks. Changing default passwords and treating unexpected links with suspicion won’t stop every attack, but both close doors that AI-assisted attackers are happy to walk through.
Source: nature.com