Security audit

Skill Compiler

Security checks for vulnerabilities and agentic risk

Overview

This is a documented skill-building workflow with some quality and scoping caveats, but no evidence of hidden, destructive, or deceptive behavior.

Installers should treat this as a capable compiler-style skill: use it on files, URLs, repositories, and media you are comfortable having the agent parse, and review generated skills before installing them. For non-Chinese video or multilingual inputs, adjust the transcription language setting or request automatic language detection.

SkillSpector

Vulnerability Patterns

Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (7)

Vague Triggers

Medium

Confidence: 86% confidence
Finding: The profile explicitly permits non-exhaustive trigger lists and states that activation uses semantic matching plus keywords, which broadens when the skill may be invoked. In a skill-compiler context, over-broad activation can cause the agent to engage on loosely related prompts or untrusted source-conversion tasks, increasing the chance of unintended execution paths, prompt capture, or misuse outside the intended boundary.

Vague Triggers

Medium

Confidence: 90% confidence
Finding: The trigger description template uses generic placeholders such as '触发词1/2/3' without requiring disambiguation, scope boundaries, or negative triggers. In a skill compiler context, this can cause generated skills to match overly broad phrases, leading to accidental invocation, prompt-routing mistakes, or unsafe activation in unintended contexts.

Natural-Language Policy Violations

Medium

Confidence: 90% confidence
Finding: This file is written entirely in Chinese and prescribes Chinese-language instructions and output structures without providing a user-language negotiation mechanism or opt-in. In a skill compiler context, this can cause downstream skills or analyses to ignore the user's preferred language, leading to misuse, misunderstanding of boundaries, and incorrect security or product decisions when users cannot reliably review outputs.

Vague Triggers

Medium

Confidence: 88% confidence
Finding: The activation conditions are broad and loosely scoped, especially triggers like long source length, repeated knowledge, or explicit user request, which can cause the optimization pass to run in contexts where it was not semantically intended. In a skill compiler, unintended execution of an optimization pass can reshape prompts, split files, or alter IR structure in ways that affect downstream safety, correctness, and auditability.

Natural-Language Policy Violations

Medium

Confidence: 89% confidence
Finding: The video transcription command hard-codes `--language zh`, which can mis-transcribe non-Chinese or multilingual media and silently distort extracted content. In this skill, transcription output becomes structured evidence for later passes, so a forced language setting can poison downstream reasoning, provenance quality, and user-visible outputs.

Vague Triggers

Medium

Confidence: 88% confidence
Finding: The trigger list uses very broad keywords such as "api", "tool", "document", and "request", which can match many ordinary prompts and cause plugin-discovery logic to activate when the user did not intend external capability use. In a skill compiler context, this can lead to over-permissive skill generation, unnecessary tool wiring, or accidental expansion of the skill's external-access surface.

Vague Triggers

Medium

Confidence: 88% confidence
Finding: The template’s description field explicitly encourages broad trigger phrases but does not require authors to define clear boundaries, exclusions, or disambiguation criteria. In a skill compiler context, this can cause the generated skill to activate in unintended situations, increasing the chance of overbroad delegation, misuse of the wrong skill, or unsafe handling of inputs outside the skill’s intended scope.

Static analysis

No suspicious patterns detected.