Classifierbert-tiny-injection-detectorms int8 ONNX, 4.5 MB, ONNX Runtime Web on WASM Embeddingsall-MiniLM-L6-v2ms int8, 23 MB, Transformers.js 4.3 on WASM

Starting

Try your own text

Scored live in this tab: a heat bar, highlights and its three nearest neighbours on the radar.

Injection radar

An agent's untrusted inputs stream past a real injection classifier and land on a map of meaning.

bert-tiny-injection-detector and all-MiniLM-L6-v2, both Apache-2.0
ONNX Runtime Web and Transformers.js 4.3; PCA computed in the page
Highlights: leave-one-sentence-out attribution

Snapshot from Node.js while the models load.

The 31 sample texts were written for this demo, so 12 of 13 is not a benchmark: on injections aimed at agents in general, a model this small misses far more often. Small models misfire both ways, so this is one layer: keep least-privilege tools and approvals.

Back to gallery