Audit Shrimp 3: Security Vetting for AI Agent Skills

The Problem

Third-party AI agent skills combine prompts, scripts, and external tool calls. Installing an unknown SKILL.md can expose credentials, read local data, or execute system-level actions. Audit Shrimp 3 treats “vet before install” as a mandatory security workflow, giving engineers a repeatable static review process before introducing unfamiliar skills.

How It Works

The skill provides a structured vetting protocol rather than generic security advice:

• Source check: evaluate whether the skill comes from an official source, a high-star repository, a known author, or an unknown origin, then adjust scrutiny accordingly.
• Mandatory code review: read all files in the skill and look for concrete red flags instead of relying on descriptions alone.
• Permission scope assessment: inspect whether the skill touches file operations, browsers, APIs, credentials, trading, or system capabilities.
• Risk classification: map the skill into low, medium, high, or extreme risk tiers, with corresponding actions such as basic review, full code review, human approval, or do-not-install.
• Vetting report: produce a structured summary that makes the decision easier to audit and revisit.

Boundaries and Caveats

It is intended for pre-installation review of AI agent skills shared through ClawdHub, GitHub repositories, or other agents. It relies on the model actually reading the skill files and executing the checks; it does not replace runtime sandboxes, dependency scanning, credential management, or organizational security policy. When a skill involves credentials, trading, system permissions, or security configuration, human judgment is required. For ambiguous sources or unclear behavior, the default action is not to install.