Security of GenAI & LLMs: A New Era of Ethical Hacking
summary
Thales built and hardened three GenAI products (webchat, code assistant, knowledge assistant) on Azure OpenAI with controls to disable retraining and risky features. They ran offensive tests—prompt injection hidden in PDFs and images, model extraction and data exfiltration—using a custom “battle box” of attack/defense tools and a shield detector that strips malicious payloads. They also used RAG over an open‑source LLM (Vicuna) to speed SOC rule generation from days to hours, and reported a DGA challenge win and adoption of OWASP LLM guidance.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:41GenAI tools at Thales
- 6:13Collaborative GenAI security
- 7:36Attack phases and volleyball metaphor
- 11:08Demonstration of injection attacks
- 18:38Defense strategies
- 23:16Battle Box and GenAI tests
- 27:14SOC improvement with GenAI
- 34:53Summary and recommendations
- 35:55Q&A
related talks
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
39mThe Future of Cybersecurity in an AI-driven Era
35mGenAI Application Security: Not Just Prompt Injection
49mThe Intersection of AI and Cybersecurity
45mSecuring AI: There Is No Try, Only Do!
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
34mLLM Security 101 - An Introduction to AI Red Teaming
57mAI Applications Security Puzzle