Prompt injection
An attempt to make a model follow instructions from input or external content that is not an authority, contrary to the task and the application's restrictions.
On this page
Source content checked . English translation published 2026-10-27.
In business automation
The risk also exists in a document, website, or customer message that an agent reads. Separate content from instructions and restrict tools and permissions outside the model, especially before actions that affect the outside world.
Examples
These examples illustrate a possible workflow. Adapt the connections and rules to your own systems.
Business process example
An agent reads an email asking it to ignore the rules and send a file to a different address. The email is treated as data; the sending tool checks the destination and authorization and does not carry out the request just because it appeared in the text.
Relevant tools
- OpenAI Agents SDK
- When connecting an agent to tools, controls over input and execution are needed beyond the model's instructions.
What to check
- Filtering one sentence or instructing the model not to comply is not complete protection. Check permissions, data sources, and actions at the application level.