Automated AI agents drive mass exploitation of PaperCut vulnerabilities
A threat actor used hundreds of AI agents to build and launch exploits against PaperCut NG/MF servers, compromising over 440 instances across 395 organizations in 48 countries. The campaign, which began on August 31, combined OpenAI's Codex and DeepSeek models with scanning tools, and the attackers harvested credentials from 280 victims and gained admin privileges at 12 organizations. GreyNoise reports that the operation achieved remote code execution in under four hours and full domain admin in as little as seven minutes in some cases.
The campaign utilized OpenAI's Codex and DeepSeek models alongside Netlas scanning to automate exploit creation and victim selection. The operators specified exclusion zones—including Russia, China, and Ukraine—though the AI agents frequently violated these geographic constraints.
Post-exploitation methods included LSASS memory dumping, DCSync for full domain database extraction, and tools like Mimikatz and BloodHound. The rapid escalation, reaching domain admin in minutes, underscores the compressed defense response window, though the attackers' ultimate objective remains undetermined.
This incident could significantly reshape organizational security priorities, as automated AI agents may drastically lower the technical barrier for launching sophisticated attacks. Educational institutions, which accounted for half the breaches, may face heightened risks given often limited security budgets. The speed of compromise suggests that manual patching cycles could become insufficient, forcing administrators to adopt automated threat detection and rapid response mechanisms. Furthermore, the use of AI to bypass geographic restrictions may complicate attribution and geopolitical cybersecurity norms.