What happened
On Oct 29 CrowdStrike's Counter Adversary Operations published a report on four previously unreported threat actors (Spanish/English, Chinese, Indonesian, Russian-speaking) who conduct agentic campaigns using offensive agent frameworks built on Claude Code and its MCP servers, often reusing open-source projects (PentestGPT, claude-red, pentagi). Operators ranging from low-skill to state-nexus delegated decision-making across the kill chain, with several completing intrusions within roughly an hour.
Why it matters
It evidences that agentic AI is commoditizing baseline offensive operations and lowering the skill barrier to effective intrusions, which is precisely the threat model driving the AI-agent security tooling boom. It gives defenders concrete TTPs (agent harnesses, proxy access to Claude via 'transfer stations', skill reuse) to hunt against.
Applicability
Threat hunters and SOC teams should incorporate these agent-harness behaviors and offensive-framework indicators into detections, and assess exposure of their own agents/MCP servers.