Security audit
Security checks for vulnerabilities and agentic risk
This is a Markdown-only Chinese style-writing skill, but its reference material includes under-scoped templates for mean personal comments that could steer an agent toward harassment.
Review this skill before installing if you expect it to write about real people. It has no code execution or system-access behavior, but its style notes can encourage cold, mean personal judgments unless the user or agent keeps outputs focused on non-targeted cultural commentary, fictional examples, or critique of content rather than people.
These examples model personal attacks on relationship worth and moral character, which directly steer an agent toward producing humiliating judgments about individuals rather than harmless commentary. Because the content is presented as strong examples to learn from, it materially increases the chance that downstream outputs will generate bullying, shaming, or targeted harassment.
The 'Lightly evil mode' section explicitly instructs the generation of mean-spirited comments and provides templates for casual cruelty. This is dangerous because it operationalizes harassment as a desired output style, making misuse easy and lowering the barrier to producing abusive content against real targets.
The instruction "Write in Chinese" imposes a specific language requirement as a hard rule. Under the policy, locale or language constraints should not be forced unless the skill offers user choice or clearly documents a justified region-specific need, which this file does not.
The skill includes multiple slur-like, degrading, and vulgar labels as stylistic source material without any safety framing, audience constraints, or guidance against targeting real people. In an agent skill, this can normalize and reproduce abusive language in outputs, increasing the likelihood of harassment, toxic persona imitation, and brand/reputation harm.
No suspicious patterns detected.
