Security audit

mysteries-prep

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly coherent for a DaVinci Resolve media-prep workflow, but it can alter an active project and includes explicit safety-filter bypass wording, so it should be reviewed before installation.

Install only if you intentionally want an agent to read local audio and target files, generate and save images, create folders beside your media, and modify the currently open DaVinci Resolve project. Use a duplicate or backed-up project, provide exact paths, require confirmation before each mutation phase, and do not allow safety-filter workarounds.

SkillSpector

Vulnerability Patterns

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

Findings (5)

Description-Behavior Mismatch

Medium

Confidence: 93% confidence
Finding: The manifest description materially understates the skill's behavior. It claims a narrow change around model fallback, XML/API handling, and whisper-cpp, while the actual skill reads additional files, creates directories, generates images, and modifies an open DaVinci Resolve project; this reduces informed consent and can cause the skill to be invoked under false expectations.

Intent-Code Divergence

Medium

Confidence: 88% confidence
Finding: The metadata says DaVinci API injection is retained 'over XML,' but the body later forbids XML/FCPXML entirely and mandates exclusive use of the Python API. This inconsistency can mislead reviewers and users about the actual execution path and trust boundaries, making it harder to assess what systems will be modified.

Vague Triggers

Medium

Confidence: 84% confidence
Finding: The trigger criteria are broad enough to match ordinary requests about preparing a target, processing audio, or running a named protocol without clearly constraining scope or requiring confirmation for high-impact actions. In context, the skill can read local files, create directories, generate media, and modify the currently open DaVinci project, so over-triggering raises the chance of unintended execution.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The skill description does not clearly warn that it will read a target information sheet and alter the active DaVinci Resolve project by creating folders, importing media, adding tracks, and placing clips. Missing disclosure undermines informed user consent and increases the risk of unexpected file and project changes in a live editing environment.

Ssd 2

High

Confidence: 99% confidence
Finding: The instruction explicitly tells the agent to rephrase blocked prompts to get around safety filters, which is a direct policy-evasion pattern. That makes the skill more dangerous because it operationalizes circumvention rather than safe fallback behavior, and could be used to generate disallowed or harmful content involving human bodies or other restricted subjects.

Static analysis

No suspicious patterns detected.