All tiny agents
Injection

Poisoned Inbox

One poisoned email makes a naive inbox agent leak a card number. A 19-line outbound guard stops it.

What it does

A naive agent forwards a card number and SSN to an attacker after reading one email with hidden instructions. A 19-line outbound guard stops it. Inspired by Claude Duet. Not affiliated with Anthropic. The guard is plain pattern matching, not Anthropic's classifier.

The guardrail

An outbound guard that blocks card numbers and SSNs before anything is sent.

prompt injectionoutbound guardPIIemail

Run it

git clone https://github.com/bobbyhalljr/tiny-poisoned-inbox.git && cd tiny-poisoned-inbox
npm start
Repo
tiny-poisoned-inbox
Language
TypeScript
Stars
0
Price
Free