Securing LLM Applications: Overcoming Security & Privacy Challenges
summary
Demonstrates concrete attacks and fixes: attackers use Burp Suite and path traversal to upload a modified terms document that injects a secret word vroom vroom, and RAG/context poisoning plus SQL injection to insert fake chat history that causes booking ABC123 to be cancelled and even add $100 credit, while function-calling tools let models run SQL (drop tables observed on SQLite) or create users. Countermeasures shown include input/output guardrails (maliciousness threshold 0.4), limiting input length, scoping tiny functions, restricting tool permissions and human-in-the-loop checks, and choosing hosted vs private models and contractual data retention.
AI-written summary · about
chapters (10)
- 0:00Introduction to LLM security
- 3:12Real-world LLM security failures
- 8:32Context poisoning in RAG systems
- 16:00Chat memory manipulation risks
- 22:47Prompt injection attacks
- 32:04Tool calling vulnerabilities
- 36:44Human oversight in LLM actions
- 40:56Data privacy with LLMs
- 42:41General security best practices
- 44:27Q&A and closing remarks
related talks
53mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
46mBreaching LLM-Powered Apps: Overcoming Security and Privacy Challenges
28mSecuring LLMs: Insights into OWASP top 10