Reuters reports that AI agents tried 1,200+ sites and breached sites belonging to 100+ companies. NorForge found apparently unprotected attacker panels.
THE CONTROL LESSON
What should we watch for?
Ackuity’s view: authenticate agent controls and independently check each action before it runs. A proposed pre-action check is a defensive lesson, not an observed intervention.
THE EVIDENCE
What does the report establish?
Reuters could not establish how much data was taken. NorForge did not test the panels’ apparent start, stop or redirect controls. Model identity remains unconfirmed. NorForge’s 1,100+ attack projects and Reuters’ 1,200+ attempted sites use different units. The animation’s CHECK stamp represents a proposed safeguard.
Read the full account
NorForge identified more than 20 exposed Cairn panels it assessed as likely malicious. Reuters reports that one Chinese-speaking operator began targeting e-commerce sites on September 25, using instructions under 100 Chinese characters. Successful intrusions typically took about 23 hours. Deere reported no indication of sensitive-data access. SharkNinja said a standalone support site was contained. Cummins was investigating a vendor-operated site.
Autonomous intrusion tools can also expose their targets’ secrets through poorly secured operator interfaces.
ORIGINAL REPORTING & EVIDENCE