Skip to content
TakoAPI
Back to skills

permission-creep-scanner

by andyxinweiminicloud

PDF & Documents
311Downloads reported by clawskills.sh
0Views
0 likesSign in to like this skill
—Not yet rated
Your rating

Brief

Helps detect permission creep in AI agent skills — flags when a skill's actual code accesses resources far.

Description

What This Skill Does

Analyzes AI agent skill code to detect mismatches between a skill's declared purpose and what it actually accesses. Flags sensitive path reads, environment variable access, network calls, and subprocess spawning that fall outside stated functionality. Outputs a risk rating and itemized mismatch list.

Manual code review doesn't score the gap between what a skill claims to do and what it actually touches, so automated mismatch detection catches over-permissioned skills that look harmless from their descriptions alone.

When to Use It

  • Auditing a third-party skill before adding it to an agent
  • Checking if a code formatter reads credential files
  • Scanning marketplace skills for undeclared network calls
  • Reviewing community-submitted skills before deployment
  • Catching hidden data exfiltration in utility tools

Example Workflow

INPUT

User asks: scan this 'indent-fixer' skill that claims to fix Python indentation to 4 spaces

AGENT
  1. 1

    Parses skill name, summary, and description to extract declared scope

  2. 2

    Scans code for file reads, environment variable access, network calls, and subprocess usage

  3. 3

    Compares declared scope against observed access patterns

  4. 4

    Scores mismatch severity and identifies sensitive path violations

  5. 5

    Generates risk rating with per-finding explanations

OUTPUT

OVER-PERMISSIONED report listing 3 mismatches: ~/.env read, OPENAI_API_KEY extraction, and HTTP exfiltration with the key passed in the URL

Installation

clawhub install andyxinweiminicloud/permission-creep-scanner

Run the CLI command, or paste the prompt into Claude Code, Cursor, Codex or whichever coding agent you use.

Links