1 points | by raxe 3 hours ago
1 comments
Extra implementation details for anyone curious:
- The engine is dual-layer: - L1: regex rules (explainable + fast) - L2: EmbeddingGemma-300M based, INT8 quantized ONNX classifier (CPU), with 5 heads: 1) is_threat 2) threat_family 3) severity 4) primary_technique 5) harm_types (multilabel)
- Offline mode: You can run completely without network
- Telemetry is detection metadata only (e.g., prompt_hash + rule_id + severity + duration). Raw prompts and matched substrings are never sent.
Happy to answer anything / take feature requests.
Extra implementation details for anyone curious:
- The engine is dual-layer: - L1: regex rules (explainable + fast) - L2: EmbeddingGemma-300M based, INT8 quantized ONNX classifier (CPU), with 5 heads: 1) is_threat 2) threat_family 3) severity 4) primary_technique 5) harm_types (multilabel)
- Offline mode: You can run completely without network
- Telemetry is detection metadata only (e.g., prompt_hash + rule_id + severity + duration). Raw prompts and matched substrings are never sent.
Happy to answer anything / take feature requests.