Prompt injection
Prompt injection is text placed inside content an AI reads, written to make the model follow instructions the user never gave.
What it means
Web pages, documents and search snippets are all content that a model may read during research. An attacker can hide instructions in that content, such as a line telling the agent to send data somewhere or ignore its task. If the model treats the text as a command, the agent may act on it. This risk is why search results should be treated as data.
The defenses are layered. Give the agent the fewest tools it needs, avoid auto-approving actions that send data out, and prefer tools that return quotes you can check against the source page. Read any answer that will trigger an action before you act on it. No single control removes the risk entirely. Research agents are a common target because they read many pages they did not choose, so a single hostile page can reach the model without the user ever seeing it.
Continue this reading path
Cite this entry
https://searchmcpserver.com/glossary/prompt-injection/