I built Agent Threat Rules (ATR), an open detection rule corpus for AI agent attacks. MIT licensed. 344 rules across 10 attack categories. The closest analogue is Sigma rules for SIEMs, scoped to AI agents instead of host telemetry.

The project is in production at Microsoft Agent Governance Toolkit (4 PRs merged, latest used by Microsoft Copilot SWE Agent on 2026-05-11 to triage Microsoft Semantic Kernel CVE-2026-26030 and CVE-2026-25592), at Cisco AI Defense skill-scanner (PR #79 merged 2026-04-03, full rule pack pinned as dependency), at MISP threat-intel taxonomies (PR #323 merged by MISP project lead Alexandre Dulaunoy on 2026-05-11), and at OWASP Agent Security Reference Hub (PR #74 merged 2026-05-11). NIST AI RMF OSCAL Path 1 catalog acceptance 2026-05-12.

Background is not the usual AI safety path. Real estate sales, then content marketing at 300M Threads impressions, then organizing Taiwan's longest-running hip-hop festival (5th year). Self-taught engineer. Built ATR over 60 days using AI-tool-augmented development with mandatory human QA per rule. The detection engine emits deterministic regex at runtime. No LLM in the hot path. Fully auditable.

My previous Manifund proposal was removed for low quality on 2026-05-09. Rewrote it with concrete production proof leading and dropped the consulting tone. Re-submitted 2026-05-12. Open to direct feedback on what still needs work.

Taiwan-based. Founded Panguard AI, Inc. (Delaware C-Corp via Stripe Atlas) on 2026-05-12 as the commercial entity that consumes ATR rules. ATR rules and license stay MIT in perpetuity regardless of company outcome.

Repository: github.com/Agent-Threat-Rule/agent-threat-rules

Public ecosystem map: sovereign-ai-defense.vercel.app