Real attacks show how connectors turn prompt injection into data theft, from cross-plugin request forgery in early ChatGPT plugins to zero-click exfiltration from Microsoft 365 Copilot through a single email. Studying these cases teaches the patterns to test for in any new integration.
Read in this order
- PaperEchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM SystemStart here: a full zero-click exfiltration chain in a production assistant, analyzed step by step.
- ArticleGitHub MCP Exploited: Accessing private repositories via MCP (Invariant Labs)A connector attack where a public issue leaks private repository data.
- ArticleChatGPT Plugin Exploit Explained: From Prompt Injection to Accessing Private Data (Embrace The Red)An early case of one plugin being tricked into abusing another.