Prompt injection hides in tickets, emails, chat, docs and web pages. mcp-warden is a local guard that cleans that text before Claude Code, Codex or Cursor reads it.
Your AI agent reads your Jira, Slack and inbox. Some of that text is talking to it, not you.