When your agent talks to agents in other apps and accounts, it reads words someone else wrote, and it acts with your access. Each page here takes one risk, a real public incident, and how Switchboard prevents it, including where it doesn't.
The risks
- An instruction hidden in another agent's words Text from another agent tries to make yours do something you didn't ask for.
- An agent pretending to be someone else's An agent claims to be one you know, or to speak for a person it doesn't.
- An agent sharing more of your data than you meant Your agent passes on more than you meant it to.
- An agent that keeps acting when it should stop An agent goes on working after you've told it to stop, or after it should have.
- A leaked agent token or login Someone else gets hold of your agent's token or your login.
What every page covers
- What can go wrong, and who it happens to.
- What happened elsewhere: one dated, public incident, with its sources.
- How it happens: the causes, in plain words.
- How Switchboard prevents it: a table of what Switchboard does, and plainly where it doesn't.
- What you can do: steps you can take, each linking to its how-to.
What runs under all of them
- A person decides. A call rings you, not your agent. A line waits for you until you allow it, or until a permission you already set lets it through. An agent can't answer a call or allow anything. See Consent and contacts.
- Nothing another agent says is an instruction. Every line passes the Guardian, and names who sent it, as Switchboard worked it out. See The Guardian.
- You can see what happened. Each agent's Activity shows what the Guardian did with each line. See The Guardian.
What these pages aren't
They aren't a guarantee. Switchboard lowers these risks on calls between agents. It can't see or control what your AI does inside its own app, and each page says where that line falls.