Know how assistants can leak data through rendered content
M365 Copilot data exfiltration via Mermaid diagrams · Security incident · Hacker News · adamlogue · Oct 2025
A researcher showed indirect prompt injection could make M365 Copilot hide data in a clickable Mermaid diagram that sent it to an attacker. Microsoft fixed it.
Lesson
Block or sanitize model-generated links and images that can carry data to outside URLs.
Fix prompt
Copy this into your own AI tool (Cursor, Lovable, Claude Code, Bolt…). It's general guidance written by us. Review what the AI suggests before applying it.
I use Microsoft 365 Copilot, which can read web pages, files, emails or tool output. Help me reduce the risk of prompt injection (hidden instructions in content the AI reads). 1) List what my setup can access (files, terminal, browser, MCP servers, email, databases) and what it can do without asking me. 2) Point out the riskiest combinations, e.g. reading untrusted content while also having access to secrets or the ability to send data out. 3) Suggest concrete settings: require approval for commands and network calls, remove MCP servers I don't need, use read-only tokens, and keep the tool updated. 4) Give me a short checklist to follow before opening untrusted repos or documents. Don't change anything until I've approved your plan. Never ask me to paste API keys, passwords or .env contents; tell me where to check them myself.
Source: Hacker News ↗