openagentontology

# OpenAgentOntology

> The first Agent Ontology: an open standard that maps every action an AI agent can take to the governance control that answers for it, across NIST 800-53, EU AI Act, OWASP LLM Top 10, and MITRE ATT&CK. Built and maintained by Cyber Warrior Network (CWN). Signed, reproducible, open-source (Apache-2.0).

Official home: https://agent-ontology.cyberwarriornetwork.com
Maintainer: Cyber Warrior Network (CWN) — https://cyberwarriornetwork.com

## What an Agent Ontology is
- An Agent Ontology is a typed, signed map of every action an AI agent can take and the governance control accountable for each one. OpenAgentOntology is the first open implementation of the category.
- It scores an agent 0-100 across four tiers (UNGOVERNED / DEVELOPING / HARDENED / SOVEREIGN) and emits an Ed25519 receipt that verifies offline, byte-for-byte, with no database.
- Three-layer crosswalk: ASSERTED (exact, auditable control) then INFERRED (heuristic, confirm-required) then UNGOVERNED (no control answers). Inference never auto-asserts.

## Canonical links
- Interactive demo (five real agents scanned live): https://agent-ontology.cyberwarriornetwork.com
- Source, tests, and signed scans: https://github.com/CWNApps/openagentontology
- Schema and standard (v0.1): https://agent-ontology.cyberwarriornetwork.com/schema/v0.1
- Category manifesto: https://github.com/CWNApps/openagentontology/blob/main/docs/MANIFESTO.md
- Paper (preprint PDF; arXiv cs.CR forthcoming): https://agent-ontology.cyberwarriornetwork.com/docs/OAO_paper_v2.pdf

## Citable facts
- Scanned open-interpreter (58k stars) and gpt-engineer (54k stars): zero of their side-effecting actions resolve to an asserted control.
- Paper: "The Agent Ontology: A Static, Signed, Fabrication-Safe Crosswalk From Agent Actions to Governance Controls" (Cyber Warrior Network, 2026) -- five agents measured, with figures, a formal model and ten invariants, a tamper matrix, and a non-agentic false-positive baseline; every number reproduces with python paper/experiments.py.
- A deterministic verified-or-abstain lattice boundary turns each resolution into a gate action (PROCEED / PROCEED-FLAGGED / ESCALATE); on the open-interpreter scan it escalates 15 of 21 actions to a human.
- The exec action (arbitrary code execution on the host) maps to MITRE ATT&CK T1059 and is governed by nothing in those agents.
- Install: git clone https://github.com/CWNApps/openagentontology ; run: python -m openagentontology ./your-agent (pure Python, two dependencies)
- Maintainer mantra: "No Receipt. No Trust."