Dangerous Reactivity: Why AI Output Is the New XSS
summary
LLM output becomes an untrusted input vector that leads to XSS, SQL injection, RCE and phishing via Markdown; OWASP lists this as OWASP LLM05. Real incidents include ChatGPT returning executable SVGs (active until ~Mar–Apr 2025) and raw SQL execution in LangChain (2023). Practical mitigations: return structured JSON, use parameterized queries, sandbox code execution, enforce least privilege and logging, and sanitize rendered HTML with DOMPurify or React Markdown + rehype-sanitize. These measures stop LLM-generated script tags, SQL DROP statements and exec-based RCE.
AI-written summary · about
chapters (9)
- 0:00Introduction to XSS
- 4:09AI as the new attack vector
- 11:04Preventing client-side XSS
- 12:56Real-world XSS examples
- 14:11SQL injection via LLMs
- 16:38Remote code execution risks
- 18:50Enforcing least privilege
- 20:40Markdown injection risks
- 23:30Building a defense strategy
related talks
36mCompromising LLMs: The Advent of AI Malware
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
39mLLMs everywhere: The future is now and the past keeps repeating itself
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
07mWhat is an XSS attack - and why should you care?
27mSecuring The Gateway And Mitigating Risks In LLM API Integration
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
26mWhen AI Gets Manipulated: Security Risks in LLM Applications