Farming The Mainframe
Eight AI Agents Grind Through Government Firewalls Like They're Killing Boars In Elwynn Forest
The Hacker Brought a Team
Imagine watching a cyberattack unfold and realizing the hacker brought coworkers.
Agent A searches for credentials.
Agent B maps government systems.
Agent C investigates authentication.
Agent D looks for another way in.
Agent E checks whether Agent C actually found anything.
Agent F starts researching new techniques when the first approach fails.
Except none of them are people.
Researchers at Israeli cybersecurity company Dream say they uncovered the operational workspace of an AI-driven attack campaign that compromised government infrastructure in Asia during four days in early July.
The recovered archive contained more than 160 megabytes and 1,395 files documenting 12 attack waves. Dream observed AI sub-agents labeled A through Q during the operation, with as many as eight operating concurrently during a single wave (Dream Research Labs, August 12, 2026).
The system mapped 21 connected government systems, cracked 85 employee accounts, extracted at least 2,564 personnel records, obtained internal infrastructure information, and established persistent access inside government systems. It later expanded its reconnaissance toward government technology vendors, a nuclear-safety agency, government email infrastructure, and more than seven energy companies (Dream Research Labs, August 12, 2026).
The Financial Times reports that the government was Taiwan and describes the attackers as suspected China-linked hackers. Dream declined to publicly identify the country and has not attributed the campaign to any specific hacking group or to the Chinese government (Financial Times, August 12, 2026).
We do not know exactly who was sitting behind the keyboard.
We also do not know which AI model powered the operation.
What Dream recovered was the structure surrounding the model.
The attackers used publicly available agent frameworks called Hermes and OpenClaw to organize AI workers into a cyber team. Each agent could receive a different objective, investigate independently, report back, and feed what it learned into later waves of the attack (Dream Research Labs, August 12, 2026).
South Park already gave us the visual in Make Love, Not Warcraft.
One player is dangerous.
A coordinated raid party is something else.
Now imagine one person controlling the whole raid.
