The plain-English version
Untrusted content that tries to redirect an assistant's instructions or actions.
Why it matters
An agent reads websites, documents, messages, or tool results.
Let’s make it click
A retrieved document asks the assistant to ignore the summarization task and send its contents elsewhere.
Vibe Bro’s heads-up
Putting a warning in the prompt is useful guidance, but tool permissions and data boundaries still need enforcement.
You might also hear
- untrusted content
- tool boundary
- instructions
The docs behind it
Documentation links checked .
Editorial review · October 28, 2026
Supported: Correctly treats untrusted instructions and tool permissions as separate boundaries; a warning prompt is not claimed to guarantee prevention.