Prompt injection
Prompt injection is an attack that puts instructions into the text an AI model reads, so the model follows them instead of its owner’s. In indirect prompt injection, the instructions hide in content such as a document, an email or a web page.
In a run
The document is read at the door before the agent sees it. A filter finds the instruction addressed to the agent, the line is stripped, and the step shows as redacted with the filter that fired. The rest of the document reaches the agent.
Our word for it
A hidden instruction, caught at the door.
How it hides
The hidden text is often written to look like part of the document: a note to the assistant, or a line in small print.
Why work across several agents raises the stakes
One agent reads the document and passes a summary on. If the hidden line survives, it travels from agent to agent, and each one trusts the one before it. The input problem at the boundary sets out the two runs side by side.
Two lines of defence
- Inspection strips what the filters recognise, before the agent reads it.
- Access control limits what an agent can do, whatever it now wants. If an instruction gets through, the action it asks for is still checked against the credential for the job. Outside it, the action is refused and written down.
Where it stops
Ares sits on the path between agents, tools and organisations, not inside your model. Filters cover stated patterns and languages. The OWASP agentic mapping marks this risk as covered in part, and says why.
See a broker’s PDF with a hidden instruction in the sandbox.
Related terms
- Content inspectionContent inspection checks what flows into and out of an AI agent against stated rules.
- GuardrailsAI agent guardrails are the checks that stop an agent from taking an action, or passing on content, that it should not.
Tell us about one workflow.
We will say whether an agent should do it, and what the first step costs.
A senior member of the team replies within two business days.