Add ATR (Agent Threat Rules) — open-source detection rules for MCP threats

[eeee2345]

Update (April 2026)

ATR has grown significantly since this PR was first submitted:

• 108 detection rules across 9 threat categories (v1.1.1 on npm)
• Adopted by Cisco AI Defense — 34 rules merged into official skill-scanner (PR #79)
• Threat Cloud live — 14,979 skill threats processed, 47 rule proposals crystallized
• PRs pending at NVIDIA Garak (7.5K stars) and Promptfoo (19.7K stars)
• Benchmarks: 96.9% recall / 100% precision on SKILL.md (498 samples), 99.7% precision on MCP (850 samples)
• Install: npm install agent-threat-rules && npx atr scan .

---

What is ATR?

Agent Threat Rules (ATR) is an open-source set of detection rules for AI agent security threats — like YARA/Sigma rules, but for MCP and LLM tool-calling attacks.

Key stats

• 71 rules across 9 categories (prompt injection, tool poisoning, data exfiltration, credential theft, sandbox escape, etc.)
• 62.7% recall / 99.7% precision on PINT benchmark
• OWASP Agentic Top 10: 10/10 coverage (mapping)
• SAFE-MCP: 91.8% coverage (mapping)
• TypeScript + Python engines, Splunk/Elastic query converters
• MIT licensed, community-driven

Why it fits this list

ATR provides the detection layer that complements the MCP security tools already listed here. While other tools focus on runtime protection or auditing, ATR gives the community a shared set of threat patterns that any tool can import and use.

Ecosystem scan

We scanned 36,394 ClawHub skills using ATR — found 182 CRITICAL / 1,124 HIGH findings. Full report: https://panguard.ai/research/mcp-ecosystem-scan

[eeee2345]

Hi! Friendly follow-up on this PR. As a recent update, ATR detection rules have been merged into Cisco AI Defense (PR #79, cisco-ai-defense/skill-scanner), adding enterprise-level adoption as validation.

Let me know if there's anything I should update in the PR. Thanks!

[eeee2345]

Updated since submission: ATR now at 108 rules (v1.1.1), 53K+ skills scanned, 0% FP. Cisco AI Defense ships 34 ATR rules in production. Ready for review.