A new benchmark, ench{}, is introduced to evaluate whether guardrails in language-model agents honor authorization. The benchmark checks authorization decisions for correctness, usability, and resulting effect, and tracks requests across guardrail changes. Experiments show that many blocked operations are due to failed reviews, and that repairing the output format or adding a second reviewer can improve decision-making.

VEPA is a framework for verified evolving formal policy for lifelong LLM agent guardrails, enabling deterministic and adaptable safety and security decisions.

A new LLM guardrail framework, ARBITER, is proposed. It uses dual-hypothesis reasoning and multi-component supervised fine-tuning for cost-effective and transparent safety decision-making. Experiments show ARBITER outperforms existing guardrails on three safety moderation benchmarks.

Safety Sentry is a lightweight guard model for AI agents that makes a per-instance three-way routing decision (execute, ask, refuse) to prevent harm in real-world environments. It outperforms existing baselines on accuracy and safety-related recall while controlling directional error rates.