Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly a Chinese AI short-drama production helper, but it includes guidance for bypassing face restrictions and removing watermarks, so it belongs in Review.
Review before installing. Use it only for assets and likenesses you have rights to, remove or ignore the face-restriction and watermark-removal guidance, and be aware that the helper script sends your text to a local LLM service and may use GW_TOKEN for that request.
Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.
---
name: manju-script-generator
description: "漫剧工业化生产全流程工具 v5.1 — 三板分离法｜七要素镜头语言｜风格顾问系统｜故事板图像生成｜镜头控制模块库｜去AI味提示词框架｜工业级素材管理"
version: "5.1.0"
The skill gives stepwise instructions to get around face-recognition or face-consistency restrictions by escalating transformations and overlays. That guidance can be used to evade identity-safety controls, impersonate people, or bypass platform protections intended to limit misuse of facial imagery.
The documented 'progressive' workaround for facial restrictions is a concrete evasion pattern, not merely neutral creative advice. In the context of an AI media-generation skill, this is especially dangerous because it lowers the barrier to impersonation, consent violations, and bypass of built-in platform safeguards around identity handling.
The skill provides operational watermark-removal steps without any authorization, ownership, or platform-policy safeguards. This materially facilitates stripping provenance or ownership markers from media, which can enable copyright infringement, fraud, and evasion of content authenticity controls.
The watermark-removal section uses practical, action-oriented instructions that directly help users strip provenance marks from images and videos. In a content-production skill, this is more dangerous because it normalizes unauthorized reuse and can aid deception about source, ownership, or authenticity.
The v5.1 '风格顾问系统' section explicitly says the final output should only be '整体风格设定' and should not include storyboards, shot lists, or technical parameters. In the same documented skill file, subsequent v5.1 sections add a storyboard image generation system and direct video-generation prompts, which contradict that 'only output' constraint within the described behavior of the skill documentation.
The only natural-language description provided is in Chinese, with no indication that users may choose another language or locale. If this skill is intended for a broader audience, that can be interpreted as a language-policy constraint without explicit opt-in or justification.
User-supplied novel text is transmitted to an LLM service over plain HTTP, which lacks transport encryption and may expose sensitive content to interception or local traffic inspection. Even though the endpoint is loopback (127.0.0.1), this still creates a privacy and data-handling risk because users are not informed that their content is being sent to a network service, and loopback services can still be proxied, logged, or bridged in some environments.
The instruction 全程纯中文输出，无英文、无乱码、无特殊符号 imposes a fixed language policy with no option for user choice. This is a natural-language policy issue because it hard-codes locale behavior rather than offering or documenting an opt-in language preference.
The document states that novel polishing must use first-person narration and later reinforces '全文统一使用第一人称"我"叙事'. This is a hard language-style constraint imposed by policy text rather than a user-selected preference.
The '场景纯净设计系统' rules say scenes must strictly contain no human figures and no character names, using absolute language such as '严禁出现任何人影'. The provided prompt template then softens this into reusable text like '不能出现其他人，无人，纯场景' and 'no humans, empty, landscape only', which is less strict and internally inconsistent with the earlier absolute constraint.
The platform-specific guidance states '可灵Kling：中文提示词' and '即梦：中文提示词+风格词' as required usage, which is a natural-language locale constraint. Because the file does not offer a language choice or explain that the requirement is strictly due to platform limitations, this may violate language/locale neutrality policy.
This manifest description lists many generic capabilities such as novel-to-script conversion and scene extraction, but it does not define specific invocation phrases, boundaries, or exclusion conditions. In a manifest file, that lack of specificity can make activation criteria ambiguous and increase the chance of unintended invocation.
GW_TOKEN is read from the environment and, if present, attached as an Authorization header for requests. There is no visible user-facing notice, comment, or prompt explaining that the skill accesses credentials from the environment and uses them for outbound requests.
The surrounding documentation and output messages say this block checks whether English quotes are used and requires Chinese quotes, but cn_quotes is computed using the same " character count twice, so it never measures Chinese quotation marks at all. This is an intent-code divergence because the code actively reports '引号格式正确' based on a check that does not implement the stated validation.
No suspicious patterns detected.