SURAKSHA 450M is out: calibrated guard for prompts + MCP tool calls
#1
by GautamKishore - opened
450M open-weight guard by Eulogik. One forward pass gives calibrated P(prompt-injection), P(jailbreak), P(PII-leak), P(tool-risk), in English + Hinglish + Hindi. Apache-2.0, trained on a Mac mini M4 for $0.
Measured on frozen heldout n=3058: injection 0.98, tool-risk 1.0, severity 0.98, worst fitted ECE 0.022. ONNX 37/37 parity.
Code: https://github.com/eulogik/suraksha
Dataset: https://huggingface.co/datasets/eulogik/suraksha-guard-38k
Feedback welcome, especially Hinglish red-team cases.