GreyNoise's Sept 9, 2026 write-up (still the clearest public case study of agentic mass exploitation) describes a likely Russian-speaking MCA that used OpenAI's Codex harness with a DeepSeek model — not OpenAI frontier models — plus commodity offensive tools to develop and fire exploits for PaperCut NG/MF CVE-2026-81578 and CVE-2026-82078. Timeline: empty workspace to first real RCE in under four hours on Aug 31; first domain admin about two hours later; once the mass campaign launched Sept 1, at least 11 orgs hit in 26 seconds. GreyNoise attributed 440 compromised instances across 395 organizations in 48 countries (education heavy; US/UK/France/Spain/Canada top countries), with credential harvest on 280, OS/domain secrets on 147, and domain admin on 12. Agents were told to avoid 28 countries but sometimes ignored that list. Cloudflare WAF blocked at least one attempt. PaperCut's later Sep 24 bulletin (updated Oct 13) tells NG/MF customers to be on 26.0.5 / 25.0.13 / 24.1.11. IOCs include 45.142.193.132, 45.158.196.75, Administrator17, Ligolo drops under C:\ProgramData\LegitSvc, and pc-*.hiv staging paths. This is the kind of harness-orchestrated campaign Context & Harness Attacks tracks: speed and scale, not novel zero-days alone.