- A Chinese-speaking hacker leveraged autonomous AI agents to breach more than 100 corporate sites, including John Deere, Cummins, and SharkNinja.
- The entire campaign ran through exposed instances of the open-source platform Carin (Cairn) using fewer than 100 Chinese characters of prompt instructions.
- Attacks averaged roughly $50 per breached target, executing multi-step vulnerability discovery and remote code execution autonomously.
Autonomous Exploitation: AI Agents Take Over Enterprise Intrusions
Cyber threat operations have taken a dramatic leap into full autonomy. According to research released by Snorre Fagerland, principal threat researcher at cybersecurity firm NorForge, a single threat actor orchestrated successful intrusions into the web infrastructure of over 100 companies by delegating the attack chain entirely to autonomous software agents.
First reported by Reuters, the victims include agricultural equipment manufacturer Deere & Company (John Deere), appliance maker SharkNinja, and engine designer Cummins. The intrusions were staged using exposed instances of Carin (also cited as Cairn), an open-source AI problem-solving platform that inadvertently exposed its management interfaces online.
Complex Attack Chains Triggered by Minimal Prompts
The scale of the operation highlighted an unprecedented level of agent autonomy. The attacker did not write detailed manual exploit sequences. Instead, by issuing an instruction containing fewer than 100 Chinese characters, the swarm of AI agents probed more than 1,200 distinct websites as of the week ending October 18.
| Campaign Metric | Details |
|---|---|
| Total Targeted Websites | 1,200+ online domains |
| Confirmed Successful Breaches | 100+ corporate systems |
| Estimated Cost Per Breach | Approximately $50 each |
| Operator Command Input | Fewer than 100 Chinese characters |
| Target Demographics | US, Canada, Australia, and Japanese e-commerce sites |
Fagerland emphasized that the tactics deployed by these agents went far beyond typical automated scripts. In one documented instance, the AI agents systematically mapped out corporate records to discover internet-facing infrastructure, pinpointed an administrative control panel, analyzed it for unpatched flaws, and formulated an exploit capable of running remote commands. As seen in other fields where AI agents execute complex workflows without human assistance, modern models are quickly adapting to offensive cyber tactics.
Commercial Impact and Underlying Risks
While SharkNinja told Reuters that the breach was confined to a server hosting product support information, other threat actors operating on the same infrastructure pursued more destructive goals. NorForge researchers observed at least one hacker successfully stealing credit card records, while another targeted a cryptocurrency exchange.
This incident confirms that the cost barrier for advanced persistent recon has collapsed. Attackers can now execute continuous, automated probing for approximately $50 per successful target without manual intervention. Security teams must prioritize identifying exposed internal orchestration platforms, monitoring perimeter portals, and mitigating zero-touch automated reconnaissance.
Attributing this entirely to autonomous AI seems like a stretch. Finding exposed admin panels and known unpatched flaws is standard automated scanning that tools have done for years. We need to be careful not to mistake high-volume vulnerability scanners with LLM wrappers for truly novel, autonomous exploit generation.