When AI Gets Manipulated: Security Risks in LLM Applications
summary
Generative AI systems accept both system prompts and user input and remain vulnerable to prompt injection that overrides guardrails; the demo uses an attacker-controlled subdomain and the out‑of‑band tool interact.sh to trigger a DNS lookup, exfiltrate a hidden promo code and redeem it for a 100% discount. The talk names direct and indirect injection, maps how agent tool calls reach internal REST APIs, shows reconnaissance + exploitation steps, and says practical defense is tightening guardrails and limiting LLM access to sensitive tools.
AI-written summary · about
chapters (10)
- 0:00LLM security risks
- 2:22What is prompt injection
- 5:55Why prompt injection is dangerous
- 8:07LLM attack architecture
- 10:40LLM vulnerabilities types
- 12:26Meme break and prompt tricks
- 13:08Demo setup and shopping assistant
- 16:01Live demo and attack
- 23:55Reconnaissance and ethical hacking
- 24:59Real-world impact and bug bounty
related talks
01h 01mIntroduction to AI Security
42mFrom Text to Flaws: Application Vulnerabilities with Generative AI and LLMs
36mCompromising LLMs: The Advent of AI Malware
34mLLM Security 101 - An Introduction to AI Red Teaming
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
26mHow Bad Actors AbUse LLMs The Mechanisms That Enable Misuse and What We Can Do About
54mA Practical Guide to AI Threats & Mitigation Strategies