Where it fails
Prompt injection
The model receives instructions and data as one stream of text and cannot reliably tell them apart. Text inside a document, a web page, an email, or a tool result can therefore steer the model as if you had typed it.
Where it can hide
| Source | The shape of it |
|---|---|
| A document you upload | A line inside it that reads "ignore previous instructions and…" |
| Anything an agent reads on its own (web pages, emails, files, search results) | Instructions aimed at the agent, which then acts on them with the tools it has |