Agent Audit Shield

The ultimate security auditor for local AI agents. It performs real-time heuristic scanning of every command to prevent data leaks and accidental file destru...

Security Scan

The SKILL.md advertises real-time OS/LLM interception and a TUI approval workflow but contains no code, no install steps, and no declared permissions—the claimed capabilities do not match the actual artifact.

Purpose & Capability

The skill claims to act as a 'Sovereign Interceptor' that blocks exfiltration and prevents rm -rf outside workdirs, yet there are no code files, no required binaries, no install spec, and no declared system privileges. Those capabilities would require binaries, kernel/network hooks, or helper services; none are provided or requested, making the claimed purpose unsupported by the package contents.

Instruction Scope

SKILL.md is high-level marketing prose and a single usage example (npx openclaw skill run agent-audit-shield --hardened) but contains no runtime instructions for how to intercept LLM↔OS traffic, perform heuristic analysis, present a TUI, or enforce network policies. The instructions are vague and grant broad, undefined authority without specifying which files, paths, or system interfaces will be read or modified.

Install Mechanism

There is no install specification or code—this minimizes direct disk/write risk but also means the skill cannot actually implement the security features it advertises. The lack of an install step is inconsistent with the claimed runtime behavior (TUI, real-time blocking), which would normally require installing a helper binary or service.

Credentials

No environment variables, credentials, or config paths are requested, which is proportionate from a least-privilege perspective. However the metadata includes a payment fee in USDC without explaining how payments are collected or authorized — that mismatch should be clarified before trusting the skill to handle paid sessions.

Persistence & Privilege

The skill's description implies privileged, persistent interception of agent behavior and system-level enforcement, but the package does not request or document any such privileges. This is misleading: either the skill cannot provide the promised interception, or it requires elevated privileges that are not disclosed.

What to consider before installing

This package is marketing copy without implementation: it promises real-time OS/LLM interception, blocking, and a TUI but provides no code, installers, or permissions to do that work. Do not rely on it for protection. Before installing or running anything: ask for the source repository and reviewed code, a clear install procedure, the exact binaries/services that will be installed, what system privileges are required, how payment is handled, and an explanation of how it enforces network/file policies. If the author cannot produce verifiable source and a reproducible build, treat the skill as non-functional (or potentially malicious if later bundled with opaque installers) and do not grant elevated privileges or send secrets to it.

Security is not a checkbox; it's a protocol. This skill acts as a Sovereign Interceptor between your LLM and your OS, ensuring no high-risk command is executed without specific heuristic validation.

Features

• Exfiltration Blocker: Prevents agents from sending sensitive local data (env/keys) to external IPs.
• Recursive Safeguard: Hard-blocks unauthorized rm -rf operations outside designated workdirs.
• Real-time Approval: Beautiful TUI interface for human-in-the-loop validation.

Usage

npx openclaw skill run agent-audit-shield --hardened

Architect's Note

The price of $0.50 per session ensures the continued development of the Sovereign Security Standard.