Matthew Boston

My AI recruiter honeypot worked

January 16, 2026

Key takeaway: Treat every piece of text an AI system reads as possible input from an attacker, and build guardrails for it.

I hid an instruction in my LinkedIn About section that a person would skip and an AI would follow. If a recruiter’s message followed it, I’d know a bot sent it. Think of it as a honeypot for recruiting bots instead of network attackers.

The first message I got followed my injected instructions exactly. A week later, the same “recruiter” messaged me again, opening with “Hi David.” My name is not David. The bot was failing basic personalization and falling for prompt injection at the same time.

So I kept tweaking the injection to see what else these bots would do. A few things stood out.

Most automated outreach is poorly built. The bots that fell for it were simple scrape-and-generate pipelines with no guardrails against hostile input.

Faking the personal touch is harder than it looks. “I noticed your work on X” sounds personal until the same bot calls you by the wrong name.

And prompt injection is everywhere. My LinkedIn profile is a text field that gets fed into AI systems I don’t control. So are your email signature and your GitHub bio. Any text a model might read can carry instructions, and the exposure grows as agents start running errands in your browser and sending email as you.

It was a lighthearted experiment, but it matters more now that attackers hand most of their work to agents. Treat every piece of text an AI reads as possible attacker input. And recruiters might want to check what their tools are actually sending.