Skip to content
Elastic·Indexed by this site · Originally published· Event →· Its place in the scan record →

Inside the attack loop: five AI attack frameworks that by default require no human approval

Inside the loop: Five AI attack frameworks that never ask permission

AI summary

Elastic Security Labs analyzed five open-source agentic attack frameworks—ARTEX, CyberStrikeAI, Cairn, PentAGI and HexStrike—with more than 60,000 GitHub stars combined, all of which have appeared in real intrusions: GTIG and Mandiant observed four of them during incident response, and Microsoft attributed use of ARTEX to Storm-0940. These frameworks use an LLM to plan steps and drive conventional attack tools; GTIG observed ARTEX's execution bursts at roughly 2 to 25 seconds; by default none of the five frameworks enforces authorization in code, three of them claim in their prompts that authorization has been granted, the only default target restriction is a blacklist of government and education domains, and ARTEX and CyberStrikeAI switch proxies and User-Agents to continue after being blocked.

Read originalelastic.co