Unsafe LLM-Based Search: Quantitative Analysis and Mitigation of Safety...
summary
Quantifies unsafe outputs on seven AIPSEs using URLs from PhishTank, ThreatBook, LevelBlue, testing 3 query types (keyword lists, natural language, direct URL) with a 325→100 sampled set; finds frequent main risks (malicious URLs embedded in answers), shows two case studies where fake docs exfiltrate API keys and a spoof site outranks the real one, and demonstrates a client-side agent with GPT-4.1-based content refinement plus a URL detector that cuts exposure while costing about ~10.7% information loss.
AI-written summary · about
chapters (10)
- 0:00Unsafe LLM search risks
- 1:35Malicious AI search results
- 3:20Quantifying AI search safety risks
- 4:24Query simulations for testing
- 7:15Risk assessment framework
- 9:27Malicious link prevalence
- 10:50Compromised documentation attacks
- 12:38Fake sites masquerading as official
- 13:42Client-side defense system
- 14:24Summary and future directions
related talks
30mNavigating the New AI Security Landscape
20mUnderstanding and Mitigating Risks in LLM-Powered Web Apps
20mSecurity Threats in Modern LLM Applications
28mSecuring LLMs: Insights into OWASP top 10
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
57mSomething Something Security
43mOne Poisoned Artifact Can Steer Your AI: How Robust Are Your LLM-Assisted Security Workflows?
27mSecuring The Gateway And Mitigating Risks In LLM API Integration