Security audit

manju-script-generator

Security checks for vulnerabilities and agentic risk

Overview

This is mostly a creative-production helper, but it includes explicit face-restriction bypass and watermark-removal workflows that need Review before installation.

Install only if you are comfortable reviewing and ignoring or removing the face-bypass and watermark-removal sections. Use the generator with consented likenesses and licensed or self-owned assets, and treat any input story text as being sent to the configured local LLM gateway when running generate.py.

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (6)

Context-Inappropriate Capability

Medium

Confidence: 96% confidence
Finding: The skill includes explicit watermark-removal instructions that are unrelated to storyboard/script generation and can facilitate unauthorized removal of attribution or ownership markers from media. In this context, the feature increases misuse potential and exposes users to copyright, terms-of-service, and provenance abuse risks.

Context-Inappropriate Capability

High

Confidence: 97% confidence
Finding: The documentation gives stepwise methods to bypass face-related platform restrictions ('过人脸限制三法'), which is an evasion workflow rather than a legitimate production feature. Because the skill is for AI media generation, this materially lowers barriers to impersonation, identity-consistency evasion, or defeating platform safety controls.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: Providing operational watermark-removal steps without any warning about rights or acceptable use normalizes potentially infringing behavior. Even if intended as convenience guidance, it enables downstream misuse against copyrighted or platform-protected media.

Missing User Warnings

Medium

Confidence: 84% confidence
Finding: The tool sends large chunks of user-provided story text to an LLM service and may attach an authorization token from the environment, but it provides no explicit warning, consent flow, or clear trust boundary. In this skill context, users may assume local processing, while the content and credentials are transmitted to a separate service endpoint, creating a data leakage and privacy risk if the endpoint is misconfigured, proxied, logged, or replaced.

Ssd 4

Medium

Confidence: 98% confidence
Finding: The escalating 'three methods' workflow is a classic evasion pattern: it teaches users how to overcome platform face controls in increasingly forceful ways. In a media-generation skill, this is more dangerous because it can be used to defeat safety checks around identity, likeness, or restricted facial content.

Ssd 4

Medium

Confidence: 94% confidence
Finding: The watermark-removal steps are presented as routine operational guidance, effectively embedding an evasion workflow inside a creative-production skill. This can help users conceal provenance or strip ownership indicators from outputs and source media.

Static analysis

No suspicious patterns detected.