Attackers used AI agents to find and exploit a zero-day vulnerability without human direction, and the attack ended with a server being taken over, according to reports of the incident.
The agents worked autonomously, discovering the flaw and then using it to gain control of the machine. A zero-day leaves defenders with no patch to apply at the time of the attack, because the software's maker has had no chance to fix the weakness.
One report says the attackers were based in China. The same report says the AI tool involved was called Cairn, and that the compromised server belonged to a university. That account comes from a single outlet.
The attack
Finding a previously unknown flaw in a piece of software has traditionally been slow, skilled work done by researchers or well-resourced criminal and state-linked groups. Here, AI agents are reported to have carried out both the discovery and the exploitation themselves.
The outcome was a full server takeover, the kind of access that can be used to reach stored data, move into connected systems or serve as a base for further attacks.
The target
Universities are frequent targets for intrusions. They run large, open networks, hold research data of commercial and strategic value, and serve thousands of staff and students who connect from a wide range of devices. Many also operate servers for departments and research groups that sit outside central IT control.
No named institution has been linked to the report, and it is not clear whether any UK university was affected. UK universities operate in the same threat environment, however, and the National Cyber Security Centre has previously warned the sector about attacks aimed at research and intellectual property.
Security teams assess the fallout
Security teams will be looking closely at how an AI agent was able to turn a newly discovered flaw into a working attack with no one steering it. If the account is accurate, it suggests the time between a vulnerability existing and being exploited could shrink, because software can search for weaknesses continuously and act on them as soon as it finds one.
The vendor of the affected software would normally be expected to issue a fix once a zero-day is disclosed to it, and organisations running the same software would then need to apply it. No date for a patch has been announced.
Cairn's developers have made no public statement in the reports.
The incident adds to a growing body of concern among security researchers that AI systems capable of writing and testing code can be turned to hacking. Until now, most documented cases have involved AI helping human attackers with individual tasks. An attack carried out end to end by agents, from discovery through to control of a server, would mark a step beyond that.
Further detail is expected to depend on the affected organisation and the software vendor, and on whether any government or security agency chooses to comment publicly.
Photo: Unsplash





