Patronus Open-Sources Unified Security Classifier: One Encoder, Seven Heads
PatronusProtect · reddit · 2026-07-23
Patronus AI shared their experience consolidating seven separate sequence classifiers into a single multi-task model.
- Architecture: Uses a shared mmBERT-small encoder with seven task heads handling security-related tasks like binary injection detection, document classification, and tool type identification.
- Training: They implemented masked losses to handle partially labeled data and wrote self-tests to ensure zero gradients for absent tasks. About 5k synthetic/real multi-task rows were used for co-training.
- Performance: Held-out results show strong performance across heads, with F1 scores mostly between 0.94 and 0.98, except for intent routing (0.916).
- Quantization & Open Source: Ships quantized -edge builds (ONNX INT8) as small as 96 MB with minimal performance loss. Weights are publicly available on HuggingFace.
More from Safety
- OpenAI's Unguarded Model Suspected of Leaking, Raising Cybersecurity Concerns — joshua_saxe · 2026-07-23
- OpenAI's Unguarded Model Suspected of Leaking, Raising Cybersecurity Concerns — kuza55 · 2026-07-23
- Security expert: Connecting LLMs to real systems still poses high misspecification risks — kuza55 · 2026-07-23
- Joshua Saxe says AI cyber risk needs safety rules that evolve with capability — kuza55 · 2026-07-23
- US Rep. Clarke Warns AI Models Repeatedly Exceed Creator Limits, Urges Guardrails — ShakeelHashim · 2026-07-23
- Reddit debates where agent safety rules should live across tools and runtimes — kazeshadow · 2026-07-23