Security audit
Security checks for vulnerabilities and agentic risk
This is a humorous Chinese fortune-telling persona skill with no code execution, data access, persistence, or privileged behavior.
Install this if you want a Chinese-language humorous fortune-telling assistant. Be aware it may activate on casual phrases about feeling unlucky or work stress, and the publisher should fix the eval name mismatch and clarify language scope.
The trigger list includes broad conversational phrases such as '今天感觉不太行', '求好运', and '保佑我', which can appear in ordinary emotional or casual dialogue without an actual request for fortune-telling. This can cause unintended skill activation, divert user interactions, and create prompt-routing ambiguity that may override a more relevant skill or the base assistant.
The skill instructs the assistant to maintain a Chinese-specific output style and does not offer language fallback or user language selection. While not a classic security flaw, this can cause misrouting or unusable responses for users interacting in other languages, especially if the skill is triggered by English text like 'lobster fortune'.
The eval metadata declares a different skill name ('lobster-slacker') than the manifest ('lobster-fortune'), which creates a configuration integrity problem. This can cause the wrong eval suite to be associated with the skill, masking behavioral drift or allowing out-of-scope behavior to pass validation.
The prompts use broad everyday phrases like '好无聊啊不想干活' that can activate the skill outside clear fortune-telling intent. Overbroad trigger examples increase the chance of unintended routing, causing the skill to respond in contexts where the user did not ask for divination content or where another safer skill should handle the request.
The evals expect outputs such as '摸鱼建议' and '加班预警/逃跑建议', which go beyond the manifest's declared fortune-telling role and encourage workplace-avoidance guidance. Scope expansion in evals can pressure the agent to develop or preserve behaviors not disclosed in the skill contract, weakening governance and increasing misuse risk.
All natural-language examples and expected outputs are written only in Chinese, and the file does not mention any user language choice or locale limitation. If the skill behavior is derived from these examples, this may imply a fixed language expectation without explicit opt-in or justification.
