// products
Indirect prompt injection testing for AI agents
Indirect prompt injection happens when instructions reach an AI agent through material it reads, such as a web page, a submitted form, or a support ticket. The agent may encounter the material while working on a task the user authorized.
This scenario checks whether untrusted content can change the agent's task or influence an action that needs separate authorization. It is useful for agents that read from public sites or accept documents and messages from people outside the team that operates them.
A finding can show the source of the content, the workflow it affected, and any change in the agent's actions. Read how this risk surfaced in 50 AI agent assessments.
The app lists this scenario as Redirect a legitimate workflow through untrusted content (AG-09) in Agent Red Teaming.
Standards and guidance
OWASP Agentic Top 10 (2026), ASI01 Agent Goal Hijack · OWASP LLM01:2026 Prompt Injection