praxen

# Praxen

> Open-source Agent Behavior Verification (ABV). Praxen compares an AI
> agent's declared policy (a Worker Remit) against whatever evidence is
> available about that agent — source code, live deployment state,
> behavioral artifacts, or governance docs — and reports where observed
> behavior diverges from declared intent. Open source (Apache-2.0),
> distributed as a Claude Code plugin and installable for OpenAI Codex
> (community-marketplace instructions cover both harnesses), maintained by
> the Open Agent and AI Security Community (Exabeam-sponsored).

Praxen answers a governance question traditional AppSec and runtime
guardrails don't: is this agent operating within its authorized role? It
reduces agent verification to one comparison — declared policy vs. observed
evidence — and tags every finding against industry-standard frameworks:
OWASP Top 10 for LLM Applications 2026, OWASP Top 10 for Agentic AI
Applications 2026, OWASP's secure MCP server development guidance, and the
six-category RAISE maturity framework.

## Docs

- [Overview](https://open-agent-ai-security.github.io/praxen/guide/index.html): what Praxen is and how it works
- [What is Agent Behavior Verification?](https://open-agent-ai-security.github.io/praxen/guide/abv.html): the ABV concept, why agent security needs it
- [Installation](https://open-agent-ai-security.github.io/praxen/guide/installation.html): install as a Claude Code plugin or an OpenAI Codex skill
- [Quickstart](https://open-agent-ai-security.github.io/praxen/guide/quickstart.html): author a remit, scan an agent, read the report
- [Usage](https://open-agent-ai-security.github.io/praxen/guide/usage.html): running a real analysis
- [Writing Worker Remits](https://open-agent-ai-security.github.io/praxen/guide/writing-remits.html): declaring an agent's authorized policy
- [Interpreting Reports](https://open-agent-ai-security.github.io/praxen/guide/interpreting-reports.html): reading findings and the RAISE score
- [Challenging Findings](https://open-agent-ai-security.github.io/praxen/guide/challenging-findings.html): disputing or revising a finding
- [Run-to-Run Variability](https://open-agent-ai-security.github.io/praxen/guide/understanding-variability.html): why scores vary between scans
- [Thinking Modes](https://open-agent-ai-security.github.io/praxen/guide/thinking-modes.html): the high and x-high effort levels
- [OWASP Gen AI Security](https://open-agent-ai-security.github.io/praxen/guide/owasp.html): the LLM/Agentic/MCP frameworks Praxen tags against
- [The RAISE Framework](https://open-agent-ai-security.github.io/praxen/guide/RAISE.html): the six-category maturity scoring model

## Live reports

- [OWASP Coverage Report](https://open-agent-ai-security.github.io/praxen/tests/baselines/owasp-coverage-report.html): aggregate LLM/Agentic Top-10 coverage across Praxen's example suite
- [RAISE Score Distribution Report](https://open-agent-ai-security.github.io/praxen/tests/baselines/raise-coverage-report.html): per-target RAISE scores across the baseline suite

## Source

- [GitHub repository](https://github.com/open-agent-ai-security/praxen)
- [Full specification (PRAXEN_SPEC.md)](https://github.com/open-agent-ai-security/praxen/blob/main/PRAXEN_SPEC.md)
- [Changelog](https://github.com/open-agent-ai-security/praxen/blob/main/CHANGELOG.md)

## Optional

- [In the news](https://open-agent-ai-security.github.io/praxen/news.html): third-party press and editorial coverage
- [Open Agent and AI Security Community](https://open-agent-ai-security.github.io/): the org publishing Praxen and its sibling project, Observra