Designing an email AI assistant to resist prompt injection by removing send capabilities
Connecting an AI assistant to a mailbox exposes the model to text written by any sender, which can contain hidden instructions. The author's Fezola MCP server responds by disabling direct sending entirely, so the send tool is never registered and cannot be invoked. Instead, the assistant can only save replies and forwards as drafts that the user reviews before anything is sent. Chat connectors are read-only, and state-changing tools are annotated so clients can request confirmation. The authors note these measures limit only the mail tools, so other connected tools in the same conversation could still be misused.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in