Understanding Prompt Injection-Techniques,Challenges and Advanced Escalation
summary
Live demos use Quarkus and LangChain for Java against models from GPT‑3.5 up to GPT‑5.4 mini and show specific failures: Rogerbot reveals system prompts and mocked env keys via structured JSON output and combined role‑play+schema tricks, a library chatbot leaks six users’ PII through multi‑turn aggregation, one‑shot payload splitting extracts names/addresses, delimiter confusion and base64 obfuscation bypass filters, and link smuggling causes the assistant to fetch an SVG that executes a shell script — proving tools/skills and MCP servers are real escalation vectors.
AI-written summary · about
chapters (19)
- 0:00Introduction and setup
- 1:34Prompt injection basics
- 5:06LLM architecture overview
- 7:11Leaking system prompts
- 9:00Structured output attacks
- 11:26Role-playing attacks
- 13:11Combined attack techniques
- 14:39Virtualization attacks
- 15:42Multi-turn manipulation
- 19:13Payload splitting
- 21:16Encoding and obfuscation
- 24:03Delimiter confusion attacks
- 27:15Link smuggling escalation
- 32:35Hidden prompt injection
- 35:48Skill file vulnerabilities
- 37:12MCP server risks
- 39:42Guardrails and mitigation
- 43:31General mitigation strategies
- 47:26Conclusion and resources
related talks
47mUnderstanding Prompt Injection - Techniques, Challenges, and Advanced Escalation
24mSecuring Language Models Against Prompt Injection
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
51mPrompt Injection: When Hackers Befriend Your AI
01h 14mBreaking the Bot: GenAI Web App Attack Surface & Exploitation
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs