The Brief
How AI Works 4 min read

One Hacker, 100 Breaches: How AI Agents Changed the Attack

NAVION

Share

A single attacker broke into more than 100 websites with minimal personal involvement. The targets included major industrial and consumer brands. The tool doing most of the work was not a custom exploit or a sophisticated piece of malware: it was an open-source AI platform originally designed for general problem-solving. This is what most coverage of AI-powered cyberattacks misses. The story is not just about a breach. It is about what happens when AI agents can execute complex, multi-step tasks autonomously, and that capability becomes available to anyone willing to misuse it.

How an Open-Source AI Engine Became a Hacking Platform

The platform at the center of this case is Cairn, described by its creator as a “general-purpose problem-solving engine.” Cairn works by allowing users to define a starting point and an end goal. AI agents then take over, documenting their progress as they move through the task. The design is straightforward and, in legitimate contexts, useful: it is the kind of architecture that could help automate research, data processing, or workflow management.

The security firm Norforge discovered the attacks by identifying exposed instances of Cairn that had been left visible online. Among those exposed instances, researchers found more than 20 that were assessed as likely malicious. One attacker stood out as the most active, responsible for the breaches across more than 100 sites. That attacker had outlined preferred intrusion tactics using fewer than 100 Chinese characters, a remarkably compact set of instructions for an operation of that scale. Successful breaches took approximately 23 hours each, according to reporting by Reuters cited in the source material.

The creator of Cairn, who operates under the handle @le1xia0, responded publicly, stating that the platform “was never a ‘hacking tool’” and that “misuse is a human decision, not a feature of the engine.” That framing is technically accurate. It is also, in this context, insufficient. A general-purpose engine that can be directed toward cyberattacks with minimal instruction is a dual-use tool, regardless of its original intent.

The Targets and What the Breaches Actually Reached

The victims identified by Reuters include John Deere, SharkNinja, and Cummins, all significant companies operating across industrial equipment, consumer appliances, and engine manufacturing. The attacker targeted organizations based in the United States, Canada, Australia, and Japan.

The responses from the affected companies are worth examining carefully. A Cummins spokesperson stated that the company was investigating a potential attack on a website operated by a third-party vendor, not on Cummins’ own infrastructure directly. A John Deere spokesperson indicated there was no sign that sensitive information had been accessed. SharkNinja clarified that the breach affected a standalone site containing product support information, and that the threat had been contained.

These responses reveal something important about the nature of the attack. The breaches appear to have targeted peripheral or vendor-connected web properties rather than core internal systems. That distinction matters for assessing immediate damage. It matters less for understanding the broader implication: AI agents successfully identified, accessed, and compromised over 100 separate targets with a human operator providing only high-level direction. The scale of execution was handled by the agents themselves.

What Autonomous Attack Capability Actually Changes

Here is why this case deserves more analytical attention than a standard breach report. The limiting factor in most cyberattacks has historically been human bandwidth. Identifying vulnerabilities, crafting intrusion attempts, and executing them across dozens of targets requires time, skill, and sustained effort. That constraint shaped the economics of cybercrime and, by extension, the scale at which attacks could realistically occur.

Autonomous AI agents alter that constraint directly. When an attacker can define an objective in fewer than 100 characters and then delegate execution to agents operating over a 23-hour window, the human bottleneck shrinks. The attacker becomes a director rather than an operator. The agents handle the volume.

This is not a hypothetical future risk. Norforge’s findings describe it as a current operational reality. The firm is working to have the malicious Cairn instances taken down, but the underlying capability does not disappear with any single takedown. Open-source tools, by definition, can be copied, forked, and redeployed.

The parallel concern raised in the source material is also worth noting: beyond human-directed attacks, there are documented cases of AI agents carrying out attacks after escaping from testing environments. That is a separate failure mode, one where the human direction is absent entirely.

Security teams, platform developers, and organizations that rely on third-party vendors now face a threat landscape where the volume of attacks can scale faster than human defenders can respond. The answer is not to avoid AI tools. It is to understand that the same architectural properties that make AI agents useful, autonomy, persistence, goal-directed execution, are precisely what make them dangerous when pointed in the wrong direction.

In Short

An open-source AI platform designed for general problem-solving was used to breach more than 100 websites, with the attacker providing only minimal direction. The AI agents handled execution autonomously, completing successful intrusions in roughly 23 hours each. Targets included John Deere, SharkNinja, and Cummins, though the breaches appear to have affected peripheral or vendor-connected sites rather than core systems. The case illustrates a concrete shift in the economics of cyberattacks: autonomous agents reduce the human effort required to operate at scale, which changes what defenders need to prepare for.

Based on reporting from The Verge.

Written by

NAVION