GreyNoise published a detailed write-up of an AI-orchestrated campaign against self-hosted PaperCut NG/MF. Starting August 31, a likely Russian-speaking actor used hundreds of AI agents — OpenAI Codex as the harness plus a DeepSeek model (not OpenAI models for inference, per GreyNoise) and commodity offensive tooling — to develop and fire exploits for CVE-2026-81578 and CVE-2026-82078. Target lists came via Netlas. Those are GreyNoise’s figures and framing; no named APT brand.
Scale underwrite (GreyNoise only): at least 440 PaperCut instances at 395 identified organizations across 48 countries; credentials harvested from 280; OS/domain secrets from 147; domain administrator at 12. Education was ~half (204 orgs). Top countries: US 98, UK 59, France 31, Spain 31, Canada 24. Speed claims from the same report: empty workspace to first real-victim RCE in under four hours; once the campaign launched, at least 11 organizations compromised in 26 seconds; one US high school went initial access to domain admin in seven minutes.
Post-exploitation paths GreyNoise observed where domain admin landed: LSASS/registry harvest then pass-the-hash; “noPac” against unpatched CVE-2021-42278/42287; or direct Domain Admins add when PaperCut ran on a DC or as a DA service account — then DCSync for a full NTDS.DIT dump. Tooling named includes Ligolo-ng, Mimikatz, Certipy, BloodHound, Rubeus, Impacket, NetExec, plus custom Rust collectors. Objective still Unclear (access brokerage vs ransomware/theft). Cloudflare WAF blocked at least one attempt. PaperCut has been shipping emergency updates — apply NG/MF patches for those CVEs and hunt the IOCs GreyNoise published.
Underwrite sheet — sourced only: CVE-2026-81578 + CVE-2026-82078; campaign from Aug 31; 440 instances / 395 orgs / 48 countries; 280 cred harvest / 147 OS-domain secrets / 12 domain admin; education 204; Codex harness + DeepSeek; Netlas targeting; speed anecdotes as above (GreyNoise, Sep 9–10 2026 coverage). Not a company-confirmed victim count. Patch PaperCut NG/MF and treat print servers as domain-adjacent risk.