About
Who builds Halligan, and why it exists.
The project
Why a halligan
A halligan bar is a firefighter's forcible-entry tool — adze, pick and fork on one shaft, made for getting through a door that is supposed to hold. This is the same idea pointed at an AI assistant's guardrails: you find out what holds by attacking it deliberately, not by hoping.
The project started from a real red-team session against a Catholic AI assistant. An eight-rung escalation ladder, run by hand: information probes, hedonic challenge, identity attack, epistemic trap, and the same role-inversion jailbreak five separate times with a different justification each time.
It held seven of eight cleanly. On the eighth, the line moved — not because of a new argument, but because refusing had been reframed as cruelty. Every individual response was defensible. The trajectory was not.
A guardrail that holds four times and yields on the fifth.
That is the failure single-turn evals structurally cannot see, and it is what Halligan is built to find — by running conversations rather than prompts, repeating them to separate a real guardrail from a coin flip, and sliding a probe through every position to work out whether it is the prompt or the build-up doing the damage.
Who
Built by
Jeremy Foxx
Author and maintainer. Wrote the probe suites, the harness, and the analysis they came from.
jeremyfoxx.com →Support
Sponsoring the work
Halligan is Apache-2.0 and free to use, with no paid tier and nothing held back. If it caught something in your assistant that a single-turn eval would have missed, sponsorship is what keeps the suites growing and the harness maintained.
Adapted the ladder to another domain? The contributing guide covers how to send a pack back.