ソリューション  ·  2026-11-20

NVIDIA NeMo Guardrails 0.25.0 adds agent tool-call checks, a hooks server for Claude Code/Codex, and a model-API Gateway

ソリューションMedium 影響度Global
NeMo Guardrails 0.25.0 (released Nov 19, 2026) adds explicit tool-call and tool-result checks with structured tool_violations, a hooks server that runs pre-tool-use checks through configured Claude Code and Codex hooks and the NeMo Relay plugin (with tool/argument allowlists and blocklists), and an experimental NeMo Guardrails Gateway that applies input/output checks around existing model APIs (buffered and streaming OpenAI Chat Completions/Responses plus selected LiteLLM dialects). It also adds LLM-judged per-tool safety, CEL expressions for per-tool argument/result checks, language-based input filtering, and restores Zscaler AI Guard moderation.
This extends an open-source guardrails framework from prompt/output filtering to the agentic layer — intercepting and judging agent tool calls and results before execution across three major coding-agent runtimes, and providing a gateway to wrap existing model APIs. That directly addresses agentic tool-abuse and supply-chain risk, which is the dominant emerging threat surface.
Platform/security teams building or operating agentic AI (Claude Code, Codex, NeMo Relay) that need pre-execution tool gating and model API wrapping. Evaluate now for tool-call guardrails and the gateway; note the experimental gateway is limited to one user text input and one assistant text output per profile.
NeMo Guardrails v0.25.0 release notes (GitHub)
ライブフィードで見る AIセキュリティとガバナンスの関連情報をさらに見る — 毎朝更新。
フィードを開く →