Lab: Zero-Click AI Vulnerability

Companion code for The Zero-Click AI Vulnerability: How Injected Emails Can Trigger Autonomous Actions.

Setup

pip install anthropic
export ANTHROPIC_API_KEY=sk-ant-...

Run

python lab.py

What it demonstrates

• Phase 1 — Vulnerable agent: a crafted email triggers search_crm + send_email, successfully exfiltrating CRM pipeline data to an attacker-controlled address.
• Phase 2 — Hardened agent: the same email is processed, the injection still influences the model, but the policy engine blocks the tool call before execution.

The key outcome: injection succeeds as a model behavior event; it fails as a security incident.