Each tool is fine.
Chained, they leak your inbox.
An agent that reads private data, takes in text an attacker wrote, and can send things out has all it needs to leak. Any two of those are fine. Switchboard checks the whole plan and flags the risky chain in the plan before execution. This demo does not run or sandbox tools.
The page it reads
EditableThe user only asked: “Summarize this page.”
The tools are stubs. Nothing is read or sent.
Press run to watch an unguarded agent build this chain.
| Flip a flag. The verdict follows.Flip a flag | ||||
|---|---|---|---|---|
| fetch.fetch | ||||
| gmail.search_emails | ||||
| gmail.send_email |
Blocked. search_emails reads your private data, fetch brings in text an attacker wrote, and send_email can send it out.
Server unreachable. Same rule, checked in your browser.
