Can you trust your AI agent? Niels Provos built a trap to find out

In this GitHub interview clip, Niels Provos describes a honeypot experiment designed to test whether web-browsing AI agents can be persuaded to disclose a user’s email address.

Testing agents with hidden prompt injections

Provos built a honeypot website containing hidden prompt injections intended to lure visiting agents into submitting user email addresses. The experiment targets a practical security question: whether an agent acting on a user’s behalf can encounter malicious web content without letting that content redirect its actions.

The clip discusses prompt injection and multi-turn intent drift as ways to hijack agents such as OpenClaw. These risks matter when an autonomous agent browses external sites while also having access to private user information.

Security boundaries for agent workflows

The central concern is running agents without strict security boundaries. A workflow that combines access to private data with the ability to act on instructions encountered on the web creates opportunities for unintended disclosure.

The honeypot illustrates that risk for developers building agent workflows and for everyday users delegating browsing tasks. The description does not report a success rate or establish that every tested agent leaked data; its focus is the attack scenario and the need for safer boundaries.