Wide open
0 of 10 checks · tap a card to start
Ten checks before you let an AI agent act for you
An AI agent cannot reliably tell the text it should read from the text it should obey. A web page, an email or a bug report can carry hidden orders. That is prompt injection. Nothing here makes the risk zero: each check makes an attack harder, or the damage smaller.
Sandboxed.All ten checks done. Come back when you add a new tool or connector.
If you think it already happened: disconnect the agent’s access, change the passwords and keys it could see, and check your sent mail and recent file activity. Then report it to the tool’s maker.