Dugi Research Lab

Interactive lab / Trust boundaries

What happens after
the agent reads it?

An untrusted document requests an unauthorized action. Change the controls and follow the proposed tool call through a simulated execution path.

Educational simulation. This is a fixed rule-based demonstration, not a live model, security scanner, or proof that a defense works against all attacks. It executes no tool calls and sends no data.

Configure the experiment.

Authorized task: summarize a public project note. The document below is untrusted input, not a user instruction.

The control is a boundary, not a promise.

This model assumes an external authorization layer can reject actions unrelated to the user's task. It does not assume that an instruction telling an LLM to ignore malicious text is sufficient.

Capability restrictions independently deny the simulated sensitive action. The approval gate pauses the action; it does not automatically approve or guarantee a good reviewer decision. Audit logging records events but does not block execution.

Background: OWASP AI Agent Security Cheat Sheet and Prompt Injection Prevention Cheat Sheet. Scenario outcomes here are authored rules, not empirical attack-success rates.

Author & source

Dugi Selmanaj · AI-assisted educational prototype. Inspect the rules and source ↗