Would a community security middleware backed by an open rule set be useful?

We maintain Agent Threat Rules (ATR, github.com/Agent-Threat-Rule/agent-threat-rules), an open-source, MIT-licensed rule set for detecting prompt injection and other agent-facing attack patterns — 655 rules across 10 threat categories as of this writing, similar in spirit to Sigma rules for SIEM but aimed at agent/LLM content instead of log lines.

Reading the Middleware System docs, `on_model_call` and `on_acting` look like a natural place to run this kind of content check — you already get `messages: list[Msg]` before it hits the model, and the maintainers have merged community-contributed middleware before (BudgetControlMiddleware, the mem0 long-term memory middleware in v2.0.3), so this isn't a new pattern for the project.

The docs currently don't have a reference security/content-filtering middleware example. Before we spend time building one, we wanted to ask:

- Is this something the project would want, or is content-safety intentionally left to users/downstream tools?
- Is there an existing pattern (even unmerged/WIP) for this kind of check we should follow instead of inventing our own shape?
- If useful, would it belong in this repo, or is a separate community-middleware package the expected home?

Happy to put together a small PoC middleware against the rule set if there's interest — wanted to check first rather than show up with a large PR.

Start in the Middleware System docs, especially the sections mentioning `on_model_call` and `on_acting`, since those are the only integration points named. Then inspect prior middleware examples (BudgetControlMiddleware, mem0 long-term memory middleware in v2.0.3) to infer the expected middleware shape and registration style. No file or test names are provided in the issue; success is a maintainer-approved decision on whether this belongs here, plus either a PoC middleware path plus docs example or clear guidance to keep it external.

由索引模型根據 Issue 內容生成。