Prompt Injection: When Hackers Befriend Your AI
summary
Demonstrates concrete exploits: direct and indirect prompt injection lets attacker-controlled text change a model’s behavior (including stored web content and plugin results), a simple jailbreak via base64 pretext bypasses system instructions, and a malicious email payload turns an assistant into a self‑propagating email worm that forwards to all contacts. Shows vulnerable tooling like LangChain’s python eval tool running arbitrary code, Chromadb/vector embeddings extending context for attacks, and Bard/plug‑in flows exfiltrating data; recommends least‑privilege, no eval, sandboxing, and structured function calls.
AI-written summary · about
chapters (10)
- 0:00Introduction and basics
- 1:53How LLMs work
- 7:19Hallucinations and context
- 12:40Prompt engineering
- 19:17Alignment and security
- 24:49Jailbreaks
- 27:51Prompt injection attacks
- 34:49Indirect prompt injection
- 42:26Glitch tokens and vulnerabilities
- 44:12Mitigation strategies
related talks
16mHarnessing AI for Malevolence: A Hackers Tale
24mSecuring Language Models Against Prompt Injection
34mHijacking Google’s CI/CD Through Prompt Injection:The New Era of AI-Based Exploits
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
43mWords as weapons: The dark arts of Prompt Engineering
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
01h 14mBreaking the Bot: GenAI Web App Attack Surface & Exploitation
45mPrompt-Jacking: The Rise of a New Supply Chain Risk