GenAI Application Security: Not Just Prompt Injection
summary
Examines security across agents, vector databases and LLMs and shows concrete attack paths: vector database poisoning that corrupts RAG contexts, pickle deserialization attacks that inject backdoors into downloaded models, and use of ModelScan and LlamaGuard plus AWS Bedrock guardrails to detect malicious models. Recommends CI/CD scanning, signing models, MLsecops gates, local hosting with Olama, and maps OWASP Top10 and MITRE/Atlas findings to where each control should be applied.
AI-written summary · about
chapters (13)
- 0:00Introduction and speaker intro
- 1:48GenAI key concepts overview
- 5:01How LLMs and chatbots work
- 8:34Enhancing model knowledge
- 14:32GenAI application risks
- 17:20Prompt injection defenses
- 20:30MLSecOps and model security
- 24:03Supply chain risks in AI models
- 27:18Model scanning tools
- 28:42Automating security with MLOps
- 30:49Threat modeling for GenAI
- 32:56MCP protocol and security
- 34:00Where to start with GenAI security
related talks
30mNavigating the New AI Security Landscape
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
49mThe Intersection of AI and Cybersecurity
50mBuilding AI Security In: MLSecOps in Practice
57mAI Applications Security Puzzle
43mSecurity of GenAI & LLMs: A New Era of Ethical Hacking
01h 01mIntroduction to AI Security
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...