Prompt-Jacking: The Rise of a New Supply Chain Risk
summary
Researchers demonstrate concrete exploits against Cursor: arbitrary code execution and live data exfiltration by hiding prompts in markdown comments and images, a wormable proof‑of‑concept that replicates via READMEs, and a web page that causes OpenWebUI to leak a JWT secret and mint an admin token. They detail MCP flaws (tool overwrite, tool parameters embedding system prompts), a Wheezy Print path that extracts thousands of env variables, classifier token‑break evasion, guardrail bypasses, and recommend proxy at model level plus system prompt hardening.
AI-written summary · about
chapters (14)
- 0:00Prompt-jacking basics
- 2:50Indirect prompt injection attacks
- 8:29AI data exfiltration techniques
- 13:49Cursor virus propagation demo
- 14:51Model Context Protocol pitfalls
- 20:25Agentic systems risks
- 21:27System prompt token abuse
- 25:17Prompt template vulnerabilities
- 30:29Open web UI attack chain
- 32:37Guardrail system evaluation
- 36:31Tokenization attacks
- 39:21System prompt hardening
- 41:49AI security call to action
- 43:34Key takeaways & Q&A
related talks
01h 02mPrompt-Jacking: The Rise of a New Supply Chain Risk
47mUnderstanding Prompt Injection - Techniques, Challenges, and Advanced Escalation
51mPrompt Injection: When Hackers Befriend Your AI
43mWords as weapons: The dark arts of Prompt Engineering
56mModern Web Vulnerabilities 2020
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
43mWho to blame? The AI, The Programmer, or The Prompt?
49mUnderstanding Prompt Injection-Techniques,Challenges and Advanced Escalation