One attacker's AI agent breached 395 orgs through PaperCut bugs

A Russian-speaking actor ran OpenAI's Codex harness against a DeepSeek model to chain two PaperCut NG/MF vulnerabilities, CVE-2026-81578 and CVE-2026-82078, into a campaign that hit at least 440 instances across 395 organizations in 48 countries. GreyNoise tracked the intrusions and found remote code execution landing in under four hours, with domain admin access following about two hours after that, per Help Net Security's writeup.

The speed is what stands out. Eleven organizations were compromised in 26 seconds. One US high school went from initial access to full domain admin in seven minutes. Education took the heaviest hit, with 204 of the 395 victim organizations in that sector, likely reflecting how widely PaperCut print management sits inside school districts.

What makes this different from a scripted exploit chain is the orchestration layer. Codex is the same agent harness developers use to automate coding tasks, here directing a DeepSeek model alongside off-the-shelf offensive tooling to work through targets. Nobody had to babysit each step. The agent found a vulnerable instance, exploited it, escalated, and moved to the next target on its own timeline, not an operator's.

Security teams building their own agent-assisted detection pipelines now face the same speed on the other side. An unpatched PaperCut instance with a public-facing web interface is a machine-speed problem, not a business-hours one. Help Net Security's report doesn't say whether GreyNoise identified a specific patch gap or misconfiguration common to the 440 compromised instances, which matters if defenders want to know exactly what to check first.