herbe assist · module
Safety
An agent that acts for you needs limits you can check. Here they are in plain words, including where they stop.
The ground rules
These come from how the apps work, not from a promise about the model.
Who holds the switch
Off until someone turns it on
An admin switches AI on per app in herbe ID, and a new access starts with AI off. After that the person keeps their own say.
- A person can untick apps on the Allow screen, so the agent never reaches them
- herbe ID records which tool was called and how it ended, never the question or the answer
- An agent may make at most 60 requests a minute for a person
- Off by default
Giving someone access to an app never opens it to AI by itself.
- Checked on every request
herbe ID checks that the person may still sign in and that AI is still on, each time the agent calls.
- Disconnect at once
The agent's next request fails after a disconnect, a sign-out everywhere, or an admin switching AI off.
The manipulation guard
A second look at text that talks to the assistant
The guard checks what people write, and the text customers wrote inside app answers, for instructions meant for the assistant. It is a setting an admin switches on, and with it off, messages are not checked. Treat it as a safety net: the firm limits are your rights, the confirmation step and Approve in the browser for what cannot be undone.
- Off, Quiet or On
Quiet asks and records, but the assistant behaves as if it had not asked. Try it there before you rely on it.
- Tool choice
Picks which groups of tools the assistant gets for a request, so it does not reach for the wrong app.
- A journal
“How the AI decided” explains each decision, can be downloaded as CSV, and lets you try the guard on a text.
How to set it up
Decide who gets AI
In herbe ID, switch AI on per app and per person. Bulk buttons turn it on or off for everyone.
Pick the privacy mode
Choose Not offered, Person chooses or Always on for anonymous mode.
Try the guard in Quiet
Watch the journal for a while, then switch the guard on.
Questions about safety
Can the agent delete a record?
Not on its own word. Deleting a calendar event or a contact, or cancelling an order, cannot be undone, so it waits until you press Approve in the browser. Smaller removals that can be put back, like a line on your own worksheet, stay within your rights.
Can it see more than I can?
No. It acts as you, and the app answers with what you are allowed to see, the same as on screen.
Is the manipulation guard a guarantee?
No. It is a check that catches instructions aimed at the assistant, and it is optional. The limits that always hold are your rights, the confirmation step and Approve in the browser for what cannot be undone.
What if someone else sets up a connector that looks like mine?
The Allow screen shows where you return to, and that line is the one to trust. Press Allow only for a connection you started yourself, and press Cancel if the return address is not what you expect.
How do I cut an agent off?
Press Disconnect under Connected apps in your herbe ID account. Its very next request fails. An admin can end a person's connections too.