models/wolfDefender/wolf_defender library

Classes

WolfDefender

Detects prompt injections and jailbreaks in text with Patronus's Wolf Defender Edge (80M parameters, Apache 2.0): for example in email, documents, web pages, and tool results before an AI model reads them.

WolfDefenderBundle

The files of Wolf Defender Edge: one model and the tokenizer.

WolfDefenderResult

The classification of one text.

Constants

wolfDefenderModelFileName → const String

Patronus's int8_int4_embeddings_x86_weight_only file: int8 weights, an int4 embedding table, and float activations. Its decisions match fp32 on every CPU that fonnx tested. Patronus's other 58 MB file, with dynamic int8 activations, does not: on x86 and WebAssembly, 94% matched, and it reversed a reference decision on the ARM CPUs of CI.

wolfDefenderTokenizerFileName → const String