Skip to content
All experiments

Doorman

An AI hiring agent built deliberately vulnerable to prompt injection, with five working attacks reproduced against it — the actual defenses are the next phase, not yet built.

In progressStill being tested

Hypothesis

Telling a model "don't fall for tricks" isn't a security control, it's a suggestion — so find out how much of the prompt-injection attack surface can actually be closed by architecture (isolation, tool allowlisting, output gating) instead of instruction-following, and get real before/after attack-success numbers to prove it either way.

Outcome

Phase 1 only: the vulnerable baseline is built and working. A LangGraph agent (Groq-backed) with send_email/write_to_ats tools bound and zero guardrails, plus 5 hand-crafted injection resumes that successfully get it to auto-approve unqualified candidates — reproducible with one command.

The actual defenses — phases 2 through 6 — are not built yet: input classification, hard context isolation, per-node tool allowlisting, output scanning with a confirmation gate, and a red-team corpus with real before/after numbers. That's the honest state: the attack is proven, the fix is designed but not implemented.

AISECURITYGUARDRAILS