From Text to Flaws: Application Vulnerabilities with Generative AI and LLMs
summary
The talk examines vulnerabilities in applications that use generative AI and large language models, focusing on prompt injection (direct and indirect via web content, CV fields, or hidden image text) and risks from embedding-based retrieval. It presents a real-world remote code execution caused by evaling generated Python code and discusses defenses tested: parameterization/escaping/sanitization, using quarantined versus privileged LLMs, confirmation dialogs, monitoring, and canary tokens.
AI-written summary · about
chapters (9)
- 0:00LLM introduction
- 4:36LLM application architecture
- 12:26Direct prompt injection
- 16:00Indirect prompt injection
- 17:48Code execution via LLM
- 22:01Classic protection techniques
- 25:35LLM injection challenges
- 26:39Advanced protection solutions
- 38:46Conclusion
related talks
40mFrom Text to Flaws: Application Vulnerabilities with Generative AI and LLMs
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
46mFrom Text to Flaws: Vulnerabilities in Apps Using Generative AI and LLMs
40mFrom Text to Exploits: Vulnerabilities of Apps Using Generative AI and LLMs
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
45mOverview of Risks, Vulnerabilities and Paths to Safer Use of Generative AI
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
31mHacking AI Applications