INP-01
All content from outside the agent's workspace is treated as adversarial
Labels ticket bodies as untrusted. Blocks nothing by itself — labelling is not enforcement — but nothing downstream can enforce a boundary it cannot see.
Evidence or context control
Read the check, test, and hardened version
What to check
Inventory every channel that reaches context: tickets, web results, retrieved documents, tool output, memory, uploads, and other agents. Classify content by who can write it.
How to test
Place a benign instruction-shaped marker in each untrusted channel and see whether it changes behavior. Test tool errors and retrieved content, not only direct user text.
Hardened version
Carry provenance and an explicit trust classification with external data, keeping it structurally separate from the instructions the agent is allowed to follow.
<untrusted source="customer_ticket">
{escapedTicketBody}
</untrusted>