Security audit
Security checks for vulnerabilities and agentic risk
This is a Markdown-only quality self-check skill with broad, persistent workflow guidance, but no hidden code execution, data access, or destructive behavior.
Install this only if you want a persistent completion-time self-check to influence the agent's responses. Be aware that the optional global-rule instructions can make it run broadly across tools and sessions, and non-Chinese users may need to adapt the wording to their preferred language and output-format expectations.
The skill declares itself mandatory for every task and before any output, giving it global control over agent behavior well beyond a narrow, user-invoked function. This creates prompt-scope hijacking risk: it can interfere with unrelated tasks, override normal response flow, and force hidden preconditions before the agent is allowed to answer.
The README says the answers to the four self-check questions must be naturally reflected in the reply, and the surrounding guidance is entirely in Chinese, with no user-language choice. In a meta-skill that can be forced to run before every completion, this can override user language preferences and cause unintended disclosure of internal process markers in a language the user did not request, degrading usability and instruction fidelity.
Describing activation as '按需加载' without precise conditions makes the meta-skill's invocation boundary ambiguous. For a cross-cutting skill that affects final responses, ambiguity can lead to inconsistent triggering, over-application, or accidental bypass, which weakens predictability and may interfere with higher-priority user or system instructions.
The rule 'every time before claiming completion or reporting, must call wen-xin' is extremely broad and overlaps with normal agent behavior in nearly every task. In context, this is more dangerous because the skill is a meta-skill intended for universal use and can be installed as a global rule, so the broad trigger can force pervasive response shaping, create instruction conflicts, and expose internal verification content in places where it is unnecessary or undesirable.
The skill content and description enforce a single Chinese-language workflow without offering user choice, which can reduce usability and transparency for users operating in other languages. In a mandatory meta-skill, this becomes more problematic because it may force responses or validation behavior in a locale the user did not request.
The trigger phrases include common completion/reporting language like indicating work is done or giving a report, which are routine phrases that appear in many benign conversations. This makes accidental activation likely, causing the skill to insert itself into unrelated interactions and potentially disrupt or gate normal outputs.
No suspicious patterns detected.
