OpenRouter launches Armor — a tool-call filter for agents

On October 15, 2026, OpenRouter launched Armor in beta — a security classifier that checks every agent tool call against natural-language rules and blocks out-of-policy calls before execution, explaining the reason to the model.

Compiled by Sergey KostenchukPublished 2026-10-16Updated 2026-10-16

2026-10-16 Safety Anthropic

🛡 OpenRouter launches Armor — a tool-call filter for agents

On October 15, OpenRouter opened beta access to Armor. The classifier sits between the model and the application and checks every tool call before execution; rules are written in English, and there are three verdicts: hard-deny, soft-deny, and allow. A blocked call is not executed, and the model receives an explanation of the reason.

🌍 Agent security is shifting from regex filters to a model-agnostic classification layer: auto mode with safeguards works with any model on the platform, not just Anthropic.

👤 If agents execute shell commands or API calls, rules are set in natural language — for example, “never push to main” — and the response includes safeguard_results with a verdict; in custom agents, Armor is a plugin with the id armor in the Responses API.

Source 1: https://x.com/openrouter/status/2110838205654470840 Source 2: https://openrouter.ai/docs/guides/features/plugins/armor

Sources