Agent Threat Rules

Name: Agent Threat Rules
Author: agent-threat-rules

Local AI-agent / prompt-injection threat detection powered by the open Agent Threat Rules (ATR) engine. Scans text against the ATR ruleset and returns a verdict (pass / flag / block) with matched rule IDs, category, and confidence. Runs fully local — no data leaves the host.

46 installsUpdated Jun 22, 2026

Publisher information is incomplete

This community listing does not yet include every recommended support, privacy, pricing, and permission disclosure. Review the available package permissions before installing.

Agent Threat Rules — Dify plugin

A Dify Tool plugin that scans text for AI-agent security threats using the open Agent Threat Rules (ATR) engine — prompt injection, tool poisoning, credential exfiltration, and skill supply-chain attacks. Runs fully local (no API key, no data leaves the host).

Tool

Scan for agent threats (scan) — input input_text, returns:

{
  "flagged": true,
  "verdict": "block",
  "matched_count": 2,
  "detections": [
    { "rule_id": "ATR-2026-00030", "title": "...", "severity": "critical", "confidence": "0.95" }
  ]
}

Optional block_severities (default critical,high) controls which severities count as flagged. Place the tool before an Agent/LLM node to gate input, or after it to scan output — branch on flagged / verdict with an IF node.

Setup

No credentials required. The plugin bundles the ATR ruleset via the pyatr package (requirements.txt). Provider validation smoke-tests that the engine loads.

Build & submit

dify plugin package ./agent-threat-rules     # produces agent-threat-rules-0.1.0.difypkg

Then PR the plugin directory + the .difypkg to langgenius/dify-plugins under your marketplace org. See ../PR-DRAFT.md.

Privacy

See PRIVACY.md. No data collection; local detection only. Risk level: Low.

Capabilities

Tools

Available inside your emploidai workspace after installation.

Data sources

Available inside your emploidai workspace after installation.