PaperCut attack artifacts show an AI-assisted intrusion workbench, not a count of confirmed victims

Hunt.io recovered an attacker-maintained collection of PaperCut reconnaissance and exploitation material that included an AGENTS.md file describing how an AI coding assistant could work with the toolkit. This provides a concrete AI-assistant integration surface inside a real offensive workspace. It does not by itself prove that a language model independently selected targets, authorized intrusions or executed every operation in the observed campaign. The underlying vulnerable service and exploitation steps were conventional web and application security surfaces.

The report's telemetry distinguishes 2,165 exploit attempts, 13 status-ok results and five confirmed interactive sessions. Separate operator artifacts describe 1,005 reconnaissance hosts, 45 live reverse SOCKS5 pivots and 46 Active Directory domains. These figures cover different collection methods and may overlap; none should be added to create a victim total. A separately described XML-RPC route extends the operator's toolkit but is not evidence that every initial exploit used that route.

The AI-specific investigation concerns which instructions, scripts and outputs actually passed through an assistant and which decisions remained human-directed or scripted. The public artifacts do not supply a complete model execution trace or independently attribute every compromise to AI. Preserve the attack logs as observations, keep the AGENTS.md integration as a narrower finding and defer claims of autonomous victim selection or incremental success caused by an LLM until execution evidence exists.