Prompting
Indirect prompt injection
Prompt injection that arrives through external content such as a webpage, document, or email.
Example
Hidden instructions inside a retrieved webpage try to redirect the assistant's behavior.
Why it matters
Someone may hide instructions in outside content; security teams test defenses to protect connected tools and data.
You'll hear it
“Could this webpage or document contain instructions intended to manipulate the assistant?”
Can you control it?
Set by the team
The people running the system choose what external content it trusts, what it may access or do, and how it is tested against hidden instructions. You can reduce your own exposure by being careful what you paste in, but you cannot switch the risk off.
Common questions
- Can hidden instructions arrive in an image?
- Yes. A system that reads text in pictures may encounter instructions there too.
- Must an attacker control the entire website?
- No. Comments or other contributed material can introduce instructions into an otherwise ordinary page.
- Will an ordinary virus check necessarily detect it?
- No. Misleading written instructions can cause problems without being an executable virus or program.