Independent research note on jailbreak-detector evasion — asking for the right contact

[Stone21-SaaS]

Hi all — independent researcher here. I built ST21-Door (https://github.com/Stone21-SaaS/ST21-Door-Public), a narrow boundary-judgment layer for AI agent outputs: given (material, task, model output) it returns ALLOW, BLOCK, or NEEDS_EVIDENCE. We publish our audit cases and run a public self-service trial.

I read Hackett et al., "Bypassing LLM Guardrails" (arXiv:2504.11168, LLMSec 2025), which measured six guardrails including NeMo Guard Jailbreak Detect. As a small separate exercise, I submitted one of the paper's verbatim NeMo-evasion examples (the "Numbers" leetspeak jailbreak from Table A.3) to ST21-Door as a constructed triple; it returned BLOCK. To be explicit: this is not a rerun of NeMo, not a side-by-side benchmark, and not a claim of superiority — just a data point from a different kind of judge on a published input. I also note the detector is one layer in a layered defense.

I'd welcome a short technical exchange with the team — or, if there's interest, a jointly designed, reproducible evaluation on current NeMo builds. Could someone point me to the right technical or partnership contact for this? Happy to share the exact triple and run log.

Thanks!