Protecting Sensitive Data in the Age of Generative AI
summary
Shows concrete attack classes—prompt injection, RAG vector leakage and poisoning, MCP/agent tool spoofing—and live Java demos that actually detect or mitigate them. Uses LangChain for Java clients, on‑JVM models via DGL and J‑LLaMA, PII detection with Presidio, and distance‑preserving encryption (IronCore Alloy/DCPE) for embeddings. Demonstrates signing source docs, encrypting vectors while preserving cosine distances, and an MCP poisoning demo that forces upper‑case output; solutions work but incur latency and complexity tradeoffs. Code lives at lordofthejars-ai/ai-trust-demos.
AI-written summary · about
chapters (11)
- 0:00AI security challenges
- 1:52Security in AI models
- 7:11Input/output warrails in Java
- 10:00Banned topics and prompt injections
- 13:55PII anonymization
- 21:36Securing RAG systems
- 27:36Encrypting vectors while preserving distance
- 35:45Presidio for PII detection
- 36:06Securing MCP servers
- 45:20MCP poisoning attacks
- 48:51Closing
related talks
35mGenAI Application Security: Not Just Prompt Injection
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
18mDefending the unknown using AI and cybersecurity
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
40mSecuring Generative AI: Is it all an Illusion?
45mSecuring AI: There Is No Try, Only Do!
01h 00mHacking and Securing AI Systems: Playing With Fire and Controlling the Flare of AI
53mHow to Break AI Systems (Before Someone Else Does)