Build an Indirect Prompt Injection Defense Gate with Python: 12 Steps in 70 Minutes

Written by

in

In this Shattered tutorial, you’ll build a runnable Taint Gate using Python’s standard library. It limits agent execution of indirect prompt injections by tracking the origins of tool-returned content, enforcing parameter allowlists, and scanning outbound data for secrets. The tutorial takes about 70 minutes to complete across 12 steps and includes one benign use case and seven attack use cases. In an offline ablation experiment, all seven attacks succeeded against the undefended agent, while all were blocked by the three-layer gate; normal email could still be delivered.

Source:Shattered · Read original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *