From Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
summary
Shows concrete exploits and defenses: prompt injection can extract hidden prompt templates and secrets (an online game "gondal" has 8 escalating levels), a LongChain "rep" tool used eval and produced a real RCE, image steganography hides instructions in near‑white pixels, and RAG uses 1,500‑dim embeddings to fetch context. Presents mitigations that actually work in practice: privileged/quarantined LLM pattern, pre‑flight/judge prompts, parameterized prompts, and active monitoring with canary tokens (fake AWS key) to detect exfiltration.
AI-written summary · about
chapters (9)
- 0:00LLM vulnerabilities intro
- 5:12Building apps with LLMs
- 15:33Prompt injection risks
- 22:25Code execution attacks
- 27:49Custom GPTs privacy bypass
- 29:40Defending prompt injections
- 40:12Advanced protection
- 48:45Key takeaways mitigation
- 50:07Closing Q&A
related talks
42mFrom Text to Flaws: Application Vulnerabilities with Generative AI and LLMs
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
40mFrom Text to Flaws: Application Vulnerabilities with Generative AI and LLMs
51mPrompt Injection: When Hackers Befriend Your AI
27mSecuring The Gateway And Mitigating Risks In LLM API Integration
54mA Practical Guide to AI Threats & Mitigation Strategies
01h 14mBreaking the Bot: GenAI Web App Attack Surface & Exploitation
20mUnderstanding and Mitigating Risks in LLM-Powered Web Apps