Security audit

music-craft

Security checks for vulnerabilities and agentic risk

Overview

This music-generation skill uses local and optional cloud backends in a disclosed, purpose-aligned way, with some setup and privacy caveats users should notice.

Before installing, confirm which backend will be used, where files will be saved, whether prompts or lyrics will be sent to a cloud provider, and whether lyrics should be user-provided, auto-written, or instrumental-only. Do not submit private lyrics, proprietary prompts, private URLs, or unlicensed reference material unless you are comfortable with the selected backend processing them.

SkillSpector

Vulnerability Patterns

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

Findings (10)

Intent-Code Divergence

Medium

Confidence: 95% confidence
Finding: The skill states it has 'zero external dependencies' even though it relies on external runtimes, CLIs, APIs, local model downloads, helper scripts, and optional cloud providers. This kind of contradictory setup guidance can mislead an agent into skipping dependency, consent, or risk checks, increasing the chance of unsafe execution paths or unexpected data egress.

Vague Triggers

Medium

Confidence: 84% confidence
Finding: Telling the agent to handle vague generation requests by staying in-scope and inferring defaults broadens activation beyond clear user intent. In practice, this can cause the skill to act on underspecified prompts, trigger external tools or cloud backends prematurely, and reduce opportunities for user confirmation around data use, cost, or licensing.

Vague Triggers

Medium

Confidence: 89% confidence
Finding: The instruction 'Use this skill unless...' creates an overly broad default routing rule that can capture adjacent tasks not safely covered by the skill. Overbroad routing increases the risk of the agent invoking the skill in contexts involving external fetching, file handling, or media processing without a tight scope check first.

Natural-Language Policy Violations

Medium

Confidence: 92% confidence
Finding: The guidance says to fall back to English when the requested vocal language is unsupported, but it does not require explicit user consent before changing the output language. That can cause silent degradation of user intent, misrepresentation of the generated content, or policy/compliance issues in contexts where language choice matters.

Natural-Language Policy Violations

Medium

Confidence: 90% confidence
Finding: The cited example infers `language=en` and then asks follow-up questions in English without user opt-in, despite the request being only a style reference and not a language choice. This can cause the skill to override user locale expectations and produce outputs in the wrong language, which is a real quality/safety issue in a routing/prompting skill because downstream generation may follow the injected default rather than the user's actual preference.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The Stable Audio example sends the user's prompt to a third-party cloud API via curl, but the surrounding documentation does not clearly warn that prompt contents leave the local environment and are disclosed to Stability AI. In a skill that may process user-supplied creative text, lyrics, or proprietary material, this can cause unintended data sharing and privacy/compliance issues, especially when users might assume all backends are interchangeable or local.

Missing User Warnings

Medium

Confidence: 89% confidence
Finding: The document states that the runtime may auto-generate lyrics when explicit lyrics are not supplied, and frames user surprise as a workflow issue rather than requiring an explicit upfront warning or consent step. In a content-generation skill, silently creating user-facing text on the user's behalf can cause mismatched expectations, unintended speech/content, or policy/compliance issues, especially when users expected instrumental output or authored text only.

Missing User Warnings

Medium

Confidence: 93% confidence
Finding: The flow tells the agent to enrich requests from an image or URL before asking the user any follow-up, but it does not require disclosure that external content may be fetched or analyzed. That can cause unintended network access, leakage of user-provided links or private resources, and surprise processing of sensitive content without informed consent.

Natural-Language Policy Violations

Medium

Confidence: 82% confidence
Finding: Defaulting vocal language to the user's message language removes an explicit choice about output language and can produce unintended content, especially in multilingual contexts or when the chat language differs from the desired song language. While not a severe security flaw by itself, it is a consent and user-intent issue that may expose preferences or create unexpected outputs.

Natural-Language Policy Violations

Medium

Confidence: 86% confidence
Finding: Automatically using the user's language for lyrics, without explicit choice, can mis-handle multilingual users and generate content in a language they did not intend for the final artifact. In some settings this also reveals or entrenches inferred language preferences without confirmation, making the behavior less transparent than it should be.

Static analysis

No suspicious patterns detected.