runtime-enforcement

VEPA is a framework for verified evolution of formal guardrails for LLM agents, using LLM-proposed symbolic policy patches validated by a deterministic verifier and an evidence-aware authority layer, outperforming baselines on agent safety and security benchmarks.

Hibana 是一个 Rust crate，允许程序从一个全局编舞（choreography）执行多角色有限协议，将其投影为每个角色的紧凑程序，并在 no_std、无分配器的运行时强制每个 send/recv/route 步骤符合协议。它支持最多 256 个角色、异步操作与仿射端点，适用于托管和嵌入式目标。

This paper tests whether stacked runtime gates (deterministic rules plus LLM judges) for agent tool calls fail independently, finding that rule+judge layers compose to ~1.9-2.1 multiplication-equivalent layers while two LLM judges only compose to ~1.2-1.5 due to correlated failures.

A developer describes building a runtime governance layer for AI agents that fixes contract mutability via cryptographic sealing and an append-only version registry, supports contractless dynamically-spawned agents through programmatic declarations, and adds a credential gateway with CI linting as a third enforcement barrier.

A practitioner asks how teams enforce safety guardrails on autonomous agents at runtime rather than relying on unenforceable policy documents, covering when agents should act autonomously versus requiring human approval for production data and external actions.

Docker Captain Karan Verma explains why AI governance must be enforced at runtime rather than relying on prompts, breaking down the execution, tool, and resource boundaries that build developer confidence in autonomous agents.

This paper identifies a silent failure mode in tool-using LLM agents where policy violations occur without tool errors or agent self-reporting. The authors propose and evaluate lightweight deterministic pre-execution gates that significantly reduce such failures in the τ²-bench airline domain.

A gateway that enforces instruction/data separation at runtime to protect agent systems from prompt injection attacks.

Introduces Faramesh, an open-source runtime enforcement layer for AI agent tool calls that checks policies before actions run, offering a solution beyond observability or LLM-as-judge.

This paper introduces a runtime execution model for autonomous agents that enforces 'Reconstructive Authority'—actions are only permitted if authority can be constructed from current state. It includes dynamic dependency resolution, a halt state for uncertainty, and a recovery loop integrating drift detection.