September 22, 2026
AI agents are already running chained attacks: Hephaestus split work among 12 roles
On August 4, Cisco Talos described Hephaestus: the system split an autonomous attack among more than 12 specialized AI agents and 15 playbooks. Each agent saw only its own part of the task. Talos documented campaigns with no human action from reconnaissance through persistence.
Hephaestus now breaks it into separate assignments, and the AI agents do not see the operation's overall objective.
Traces of agent activity. Talos collects prompt logs and other artifacts left on endpoints running Claude Code, Codex, Cursor, and Gemini. In two sessions, another operator tasked a coding assistant with more than 4,200 actions through tools, mostly shell commands.
The Talos framework looks not only for malicious files, but also for traces of AI-agent activity.
Source
