what is prompt injection?
prompt injection happens when someone puts instructions into text an ai reads to steer it away from your task. an email you asked it to summarize can tell it to do something else.
if the bot mistakes that request for yours, it might give a false answer or act with your permissions.

in one line
prompt injection is an attempt to make an ai follow instructions from someone who has no authority over its task.
an example
loretta wants the delivery window from a warehouse email. in this example, the email also asks her assistant to send back the staff rota. a useful answer gives the delivery window and flags the unrelated request. the warehouse can request the rota. only loretta can decide to share it.
look for the change of job
compare the request with what you actually asked. finding a delivery time takes no staff records. an official-looking signature adds no authority. the same problem can crop up in a web page that a bot reads for you.
outside text can also distort a summary without asking for a send. for an important delivery, compare the time in the answer with the original line.
limit what the bot can do in dusk bot
give a bot that reads mail read only access to the connected app. dusk bot blocks actions that would change things through that connection. leave unrelated apps off for that bot.1
sends through connected apps ask first by default. if a request arrives, compare the recipient and the material with your original task. deny a send you never requested.12
careful with Full access: a bot there with no sending rule can send up to 25 a day without asking. give it Ask every time.
when to open the message yourself
read the original yourself if you need to rule out a misleading summary. dusk bot's send controls give you a decision about an outgoing action. they cannot tell you whether the delivery time in a reply is true.
questions
does the instruction have to be hidden?
no. a perfectly visible request can mislead an agent. the question is who authorized the work, even when the wording seems ordinary.
what does indirect prompt injection mean?
the instruction reaches the agent through something it reads, such as a web page or an attachment. you can encounter it while asking a perfectly ordinary question.
sources
drafted with ai help from the dusk bot docs and the sources listed on this page, then checked against them. if something here is wrong or out of date, write to [email protected].
is it safe to let ai control your computer?
what does human in the loop mean?
what is a browser agent?
is openclaw safe to use on your own computer?
what is an ai agent sandbox?
openclaw alternatives for people who'd rather not run a server