One Bad OSINT Can Ruin Everything: How Secure is Your CTI RAG System, and Can You Minimize the Risk?
summary
Demonstrates a practical exploit where inserting a single malicious OSINT report into a CTI corpus plus targeted retriever+generator optimization (discrete token search for attention-steering quotation marks) makes RAG systems fabricate attacker-chosen outputs; authors run a live end-to-end demo on CTI-Bench, move the injected document from rank ~15 to rank 1 and reach an 86% attack success rate. Includes working Jupyter demo, code, and low-cost mitigations: context sanitization, semantic-outlier detection, and guardrails.
AI-written summary · about
chapters (9)
- 0:00OSINT risks in CTI RAG
- 3:19RAG for CTI workflows
- 7:13Malicious doc injection in CTI
- 10:24Poisoned docs for RAG attacks
- 12:31Attention mechanism in RAG poisoning
- 17:29Live demo: RAG poisoning attack
- 21:42Boosting attacks with attention heads
- 22:46Key findings and defenses
- 25:16Q&A
related talks
43mOne Poisoned Artifact Can Steer Your AI: How Robust Are Your LLM-Assisted Security Workflows?
26mBreaking RAG Systems: Exploiting Vulnerabilities & Hardening Your GenAI Applicatio...
16mSoK: Automated TTP Extraction from CTI Reports – Are We There Yet?
27mStandard Classification: Connecting the Dots: How Threat Intelligence Protect...
30mNavigating the New AI Security Landscape
20mAI Red Teaming and Automation: Exploring Societal Risks in...
56mAttacking AI
17mOSINT to Compromise