berabuddies/Semia

Semia, security audit for AI agent skills.

What it solves

Semia addresses the security risk of using third-party AI agent skills, which are often markdown files containing shell commands, network calls, and tool invocations. Because these skills run with the user's credentials and data on their local machine, Semia provides a way to audit them and understand exactly what a skill can do before it is trusted and executed.

How it works

Semia treats a skill file as data rather than executing it. It follows a four-stage process:

  1. Prepare: Normalizes the skill text.
  2. Synthesize: Uses an LLM to extract potential behaviors and facts from the skill.
  3. Detect: Evaluates these facts against deterministic Datalog rules to identify security violations.
  4. Report: Generates an evidence-backed report where every finding is tied to a specific line of source code.

Additionally, Semia can automatically suggest fixes for identified vulnerabilities by tracing violations back to their root cause and using an LLM to generate a patch for the skill file.

Who it’s for

It is designed for users and developers who integrate AI agent skills into their workflows (using tools like Codex, Claude Code, or OpenClaw) and need a verifiable security audit of those skills before deployment.

Highlights

  • Evidence-backed reporting: Every security finding is grounded in a specific source line.
  • Deterministic boundary: Uses Datalog rules to ensure that only checked, evidence-grounded facts are reported.
  • Automated repair: Can generate patches to fix problematic content or add security constraints.
  • Multi-format output: Supports Markdown, SARIF (for GitHub Code Scanning), and JSON.
  • Plugin support: Integrates directly as a plugin for Codex, Claude Code, and OpenClaw.

Related

  • Project
    cisco-ai-defense/skill-scanner

    A security scanner for AI Agent Skills that detects prompt injection, data exfiltration, and malicious code using static analysis, behavioral dataflow, and LLM-based semantic analysis.

  • Project
    NVIDIA/SkillSpector

    SkillSpector is an open‑source NVIDIA tool that scans AI‑agent skills (plugins, prompts, code) for security problems. It supports static analysis of 71 known vulnerability patterns, optional LLM‑based semantic checks, live CVE look‑ups, baseline suppression, and multiple output formats (terminal, JSON, Markdown, SARIF). It can run as a CLI, Docker container, or MCP server, making it useful for developers, CI pipelines, and production agents.

  • Project
    XiNian-dada/Fuck_My_Shit_Mountain

    A codebase auditing skill for AI coding agents that provides structured, evidence-based reports across 25 dimensions including security, stability, and AI safety.

  • Project
    joeseesun/qiaomu-meta-skill

    A meta-skill that transforms prompts and workflows into researched, evaluated, and securely published AI agent skill packages.

  • Project
    modiqo/skillspec

    A CLI and contract system that converts prose-based agent skills into structured, auditable specifications to ensure AI agents follow safety rules and provide proof of completion.