1 min lesson
Contain prompt injection
Trace how permissions, hooks, isolation and human review stop an instruction hidden in external content.
Step 1 of 3
A web page, Jira comment or document can contain text that tries to direct the agent. Treat external content as untrusted data.
Prompt injection happens when untrusted content changes how an agent behaves. Assume the model may follow that content, then use permissions and execution controls to limit the result.