Security audit
Security checks for vulnerabilities and agentic risk
This skill is a local writing-style checker, but it explicitly coaches users to reduce AI-detectability before sending sensitive communications.
Install only if you are comfortable with a tool whose main workflow is making text look less AI-generated. Do not use its score as proof of authorship, and be careful using it where authorship transparency is expected. The publisher should remove the unnecessary Write permission or clearly justify it.
The file is entirely written in Chinese and labels itself as the AI skill instructions, with no indication that other languages are supported or that Chinese is optional. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale constraint is explicitly justified.
The manifest description lists trigger phrases such as "AI味" and "human signal" without narrowing context or exclusion conditions. These are broad enough to appear in ordinary discussion about writing quality, which could cause unintended activation for a markdown/manifest file.
The manifest makes contradictory safety assertions: it claims the skill is read-only and has no file writing, yet explicitly permits the Write tool and filesystem access. That mismatch can mislead reviewers and users about the skill’s actual capabilities, weakening trust boundaries and making privilege creep easier to hide.
The read_when section includes broad phrases like "像AI写的", "AI味", and "de-ai" alongside usage scenarios, but it does not specify when these phrases should or should not activate the skill. Without scope constraints or negative examples, the trigger surface is ambiguous for a manifest file.
The skill is explicitly positioned to help users make AI-generated text less recognizable as AI-generated, which is an evasion-oriented objective. This can facilitate deception in contexts where authorship transparency matters, such as job applications, professional communications, or trust-sensitive submissions.
The workflow instructs users to repeatedly modify text until the score indicates it appears more human, which operationalizes iterative evasion. That turns a heuristic style checker into a tuning loop for bypassing AI-authorship suspicion and increases misuse potential.
The Chinese section mirrors the same anti-detection purpose, reinforcing that evasion is a core use case rather than incidental wording in one language. Duplicating that framing broadens accessibility of misuse and confirms the intent across the skill’s audience.
The Chinese workflow gives explicit iterative instructions to push the score below 30 before sending, which is concrete evasion coaching. This materially increases misuse risk by providing a measurable optimization target for hiding AI-like characteristics.
The module docstring, CLI description, and output labels are all presented only in Chinese. This creates a language/locale policy concern because the skill forces one language for interaction without documenting that it is region-specific or allowing the user to choose another language.
Argument descriptions, status messages, and result text are emitted exclusively in Chinese. Because the skill does not provide a user-selectable language or clearly justify a locale restriction, it violates the natural-language policy category for forced language choice.
The file presents all core naming and descriptions in fixed Chinese/English form and does not describe user language preference or opt-in behavior. While bilingual content is less severe than a single forced language, it still reflects a predetermined locale presentation rather than an explicit user-selected language mode.
No suspicious patterns detected.
