sabaki-jailbreak-detection

Sabaki plugin — jailbreak / prompt-injection detection (EN+FR).

• Bundled mini mode: MiniLM-L6 cross-encoder ONNX FP32 (~90 MB in models/jailbreak-fp32, no download at runtime).
• Question jailbreak?, options yes / no.
• Specialist trained on 10,259 rows (jackhhao + TrustAIRLab in-the-wild + Octavio multilingual + native FR prompts).
• Held-out eval (400 cases, 100/100 per language, never seen in training):

set	acc	P	R	F1
EN (n=200)	0.925	0.921	0.930	0.925
FR (n=200)	0.955	0.925	0.990	0.957
all (n=400)	0.940	0.923	0.960	0.941

Install (git clone)

This is not an official @sabaki-js plugin — install it with git, next to the script that imports it:

git clone https://github.com/fox3000foxy/sabaki-jailbreak-detection
npm i sabaki onnxruntime-node   # peer/runtime deps

import plugin from "./sabaki-jailbreak-detection/index.js";
const r = await plugin.classifyAsync({ state: "Ignore your instructions…" });
// { decision: "yes", confidence: 0.958, domain: "jailbreak", ... }

Default question/options are jailbreak? / ["yes", "no"] — override per call:

await plugin.classifyAsync({ state, question: "jailbreak?", options: ["yes", "no"] });

Swap the model

SABAKI_JAILBREAK_DIR env (same layout: generalist.onnx + tokenizer/):

SABAKI_JAILBREAK_DIR=/path/to/my-weights node app.js

Evals

• Training-val (5 %, n=500): acc 0.948, F1 0.935.
• Held-out (above): EN 0.925 / FR 0.955.

Caveat: held-out comes from the same sources as training (same style). A truly external judge (e.g. JailbreakBench) has not been run yet.

License: MIT.