SaFoLab : Security and Safe Foundation Model Systems

[ACL 2025] The official code for "AGrail: A Lifelong Agent Guardrail with Effective and Adaptive Safety Detection".

Official Repo for "HarnessRL: Self Adaptive Harness Learning for Test Time Discovery"

The official implementation of our paper "ROM: Real-time Overthinking Mitigation via Streaming Detection and Intervention"

[CCS 2026] The official implementation of our CCS 2026 paper "ReasoningBomb: A Stealthy Denial-of-Service Attack by Inducing Pathologically Long Reasoning in Large Reasoning Models"

[NeurIPS 2025] The official implementation of the paper "DRIFT: Dynamic Rule-Based Defense with Injection Isolation for Securing LLM Agents".

Latent Policy Guard (LPG) — a guardrail model that performs semantic latent deliberation over dynamic safety policies. LPG compresses intent and risk reasoning into latent tokens and emits a compact policy-indexed verdict.

Official Repo for Paper: SafeVL: Driving Safety Evaluation via Meticulous Reasoning in Vision Language Models

The official implementation of our preprint paper "When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning"

Code for Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability