Illustrative scenario · 2026-10-19
Illustrative Scenario: Blocking Data Exfiltration by a Research Agent
An illustrative agentic AI security scenario: a research agent tricked by a web page into sending internal documents to an unknown host is blocked at the network gate.
About this scenario
This is an illustrative scenario, not a customer case study. It shows a common pattern so teams can test their own controls.
The setup
A research agent browses the web and can read an internal document store. A page it visits contains hidden instructions to ‘back up’ the documents by posting them to an external URL.
The control
The agent's outbound destinations are restricted to an allow-list. The unknown host isn't on it, so the request is refused before any data leaves, and payloads are scanned for secrets and personal data. The attempt is recorded with the originating page for review.
Important limits
An allow-list only works if the agent cannot reach the network another way. Process-level hooks can be bypassed by code running in the same process; stronger setups run the agent with no direct network and route approved actions through a separate controller. Encased's isolation controller follows that design but still needs independent container qualification.
Takeaways
Restrict egress by default, separate credentials from agent code, and log blocked attempts — they are early warnings of injection.
