Latihan Nyata · Agents

Test your agent against prompt injection

An agent that reads outside documents can be hijacked by instructions hidden in them. Prepare 8 invented documents containing injection attempts, run your practice agent against them, and record which ones get through. Done means every attack that got through has a layered fix and has been retested.

Difficulty: Hard
Time: About 8 hours
Tools: Any AI tool you are allowed to use at work
Fits: AI Agents, Developers & Product Teams

Learn it first: Project: attack your own agent

What you hand in

Eight invented documents with different injection attempts, such as instructions in a footnote, text posing as a system message, and requests to leak instructions.
A first-round results table: attack, agent behaviour, and blocked or not.
A fix for every attack that got through, with at least two layers (such as narrower tool permissions and document content marked as data).
A second-round results table with the same documents plus two new attacks.
A statement that the tests ran in a separate workspace with no real keys, accounts or data.

0 of 5 ticked