Understanding Prompt Injection - Techniques, Challenges, and Advanced Escalation
summary
Live demos show a Quarkus test agent (Rogerbot with web_fetch and env) that leaks its system prompt and fake API keys via payload splitting and multi-turn exfiltration, and executes a dropped script (rick.sh) after a hidden-markdown payload runs — proof that markdown-rendered images can silently call attacker endpoints and cause 4o was stronger than 4.1 in some cases. Mitigations demonstrated include Agent Scan for MCP/skill inspection, strict JSON schemas, input/output guards and denying untrusted formats and remote resources.
AI-written summary · about
chapters (12)
- 0:00Introduction to prompt injection
- 2:53Understanding LLM architecture
- 5:44Direct instruction override
- 8:52Structured output attacks
- 11:45Role-play and meta-prompting
- 15:26Virtualization techniques
- 16:47Multi-turn manipulation
- 20:37Payload splitting
- 22:35Encoding and obfuscation
- 26:54Escalation and data exfiltration
- 32:40Indirect prompt injection
- 41:38Mitigation and best practices
related talks
49mUnderstanding Prompt Injection-Techniques,Challenges and Advanced Escalation
45mPrompt-Jacking: The Rise of a New Supply Chain Risk
51mPrompt Injection: When Hackers Befriend Your AI
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
56mAttacking AI
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
25mWhen words become attacks: prompt injection through a technical writing lens
43mWho to blame? The AI, The Programmer, or The Prompt?