Comment by brianmcnulty

You should also assume the user can read any data you send back from a tool call or data you add to a user response. If any part of the input or output is controllable by an attacker, you should be assuming some prompt injection is possible that allows them to access all data and tool calls the agent had and has access to.