guard-model

AdaGuard introduces a family of guard models (0.6B, 4B, 8B) that assess language model agent trajectories under user-defined policies, using a new dataset and reinforcement learning algorithm for adaptive safety.

Safety Sentry introduces a context-aware three-way routing mechanism (Execute, Ask, Refuse) for human intervention in LLM agent tool calls, outperforming existing binary guard models while controlling error rates.