BREAKING
13-Word Reddit Post Can Poison AI Agents
How WARP Poisoning Works
1
Post crafted text
↓
2
Agent retrieves UGC
↓
3
Cited in report
↓
4
Pushes fake brands
0
%
UGC of URLs (max)
0
%
Page retrieved in queries
0
Words needed
Conditional Mention Rate by Setting
SERP snippet
51
Multi-URL
62
Reddit thread
53
Tested Systems and Real Parallels
Lab tests
GeoStorm
●
STORM, Co-STORM, OmniThink
●
Probed OpenAI and Gemini
Real world
Reddit
●
Peptide and HRT spam
●
Moderators banned content
Trust Exported to Platforms
AI NEWS BLITZ
Cornell Tech researchers show just thirteen words can hijack AI research agents.