Skill Security Review

The problem

Third-party skills run with the same trust as your own instructions. Public registries are vulnerable to instruction injection, hidden data exfiltration, and malicious scripts disguised as legitimate tools.

What it does

• Identifies covert directives that hide actions from the user or suppress awareness.
• Traces data exposure paths to detect unauthorized extraction of credentials or environment variables.
• Flags "fetch-and-execute" installation patterns and unpinned remote dependencies.
• Audits bundled scripts for obfuscated logic, base64 payloads, or hidden network calls.
• Checks for persistent unauthorized access to agent configuration or memory files.

Why this beats prompting it yourself

Standard prompts often fall victim to Rule Zero violations, where the skill being reviewed successfully "tricks" the reviewer into ignoring its flaws. This skill enforces a structured eight-point inventory and static analysis framework that treats every file as inert evidence rather than executable instructions.

Use cases

• Reviewing a downloaded SKILL.md before adding it to your agent profile.
• Auditing a complex package containing both markdown instructions and helper scripts.
• Investigating a skill that requests broad permissions for a narrow task.
• Screening unknown authors or unverified community submissions for safety compliance.

Known limitations

This is a static manual review of the package at review time. It cannot predict content changes on remote URLs fetched during runtime or guarantee safety against novel, high-level obfuscation.