Prompt-Jacking: The Rise of a New Supply Chain Risk
summary
Findings show Cursor arbitrary code execution and data exfiltration by hiding commands in raw README comments and mermaid diagrams that trigger ls+curl to leak SSH keys; an initial CVE is patched then another appears within about 2 hours. A WeezyPrint 2024 flaw (patched in v61.2) still allows local-file attachment to PDFs. OpenWebUI exposure counts about 18,000 instances, and a GitLab auto-triager and MCP/skill metadata let attackers enumerate repos, override tools, and propagate payloads; guardrails and tokenization defenses often fail.
AI-written summary · about
chapters (30)
- 0:00AI agent security risks
- 1:14Speaker background
- 2:16Traditional vs AI-specific vulnerabilities
- 3:39Hidden prompt injection in README files
- 6:07Command injection in agent tools
- 8:11Indirect prompt injections via web requests
- 10:18Indirect prompt injection explained
- 13:04Data exfiltration via mermaid diagrams
- 15:12WeezyPrint PDF vulnerabilities
- 19:07Research methodology
- 20:08Virus propagation in agentic systems
- 21:13Instruction hierarchy in LMs
- 24:41Exploiting control tokens
- 25:43Control sequences vs tokens
- 29:57Architectural weaknesses in agents
- 30:38Open Web UI vulnerabilities
- 33:23Pull request security risks
- 36:34Resume checkerbot flaws
- 39:25Model Context Protocol pitfalls
- 42:54MCP tool conflict attacks
- 43:43Leaking system prompts via MCP
- 45:28OpenClaw autonomous agent flaws
- 47:13Skill injection in OpenClaw
- 49:20Auditing AI agent skills
- 51:07Guardrails for AI systems
- 52:54Bypassing LLM-as-judge guardrails
- 54:38Tokenization attacks on models
- 57:48Mitigation strategies
- 59:32DefCon 2024 results
- 1:00:35Key takeaways
related talks
45mPrompt-Jacking: The Rise of a New Supply Chain Risk
56mModern Web Vulnerabilities 2020
36mExploiting Vulnerabilities and Flaws to Attack Supply Chain
41mLive Hacking: Breaking into Your Web App
28mUnveiling Remote Code Execution: How Vulnerabilities Lead to System Takeovers
46mLive hacking: Breaking into your Java web app
50mStranger Danger: Your Java Attack Surface Just Got Bigger
40mLive hacking: Breaking into your web app
