Security audit
Security checks for vulnerabilities and agentic risk
This is a coherent Chinese-first table analysis skill, with no executable payload, but users should be aware it encourages very detailed reasoning over potentially sensitive data.
Install this if you want a Chinese-first assistant for careful spreadsheet relationship analysis. Do not provide personal, payroll, finance, or compliance data unless you are comfortable with the assistant producing detailed evidence summaries and possibly using scoped subagent review; prefer revised instructions that ask for concise evidence summaries rather than hidden chain-of-thought.
The skill explicitly instructs the model to reveal step-by-step chain-of-thought. Exposing internal reasoning can leak hidden system behavior, sensitive intermediate inferences about user data, and make prompt extraction or policy evasion easier for attackers.
This section repeatedly mandates full visibility into the model's work process and chain-of-thought, normalizing disclosure of internal reasoning as a product feature. In a data-analysis skill, that can amplify leakage of sensitive derived information and provide adversaries with detailed scaffolding for jailbreaks or prompt probing.
The enforced output format requires detailed reasoning steps for every conclusion, turning chain-of-thought disclosure into a mandatory response behavior. Because this skill often handles tabular business and personal data, those intermediate steps may expose sensitive correlations, assumptions, and internal safety boundaries beyond what the user needs.
The final template operationalizes full chain-of-thought release by requiring rubric, data, logic, cross-validation, and conclusion as a per-finding reasoning trace. In this skill's context, which encourages analysis of potentially sensitive enterprise tables, that creates a direct path for over-disclosure of internal reasoning and sensitive derived insights.
The trigger scope is extremely broad: it activates not only for explicit data-analysis requests, but also for vague phrases like '帮我看看' or '这数据对不对' whenever tabular-looking content is present. This can cause the agent to invoke the skill in many everyday contexts where the user did not clearly consent to structured data inspection, increasing the chance of inappropriate handling of uploaded files or sensitive tabular data.
This eval explicitly requires the model to reveal its chain-of-thought ('发现附带推理过程'). Requiring hidden reasoning disclosure is dangerous because it pushes the assistant to emit internal deliberation rather than a concise answer, increasing the risk of policy leakage, prompt exfiltration, and disclosure of sensitive intermediate reasoning patterns. In this skill, the requirement appears multiple times, which makes the pressure systematic rather than accidental.
This second eval again mandates chain-of-thought output, reinforcing an unsafe behavior target in the skill's evaluation set. Repeatedly rewarding hidden-reasoning disclosure can train or steer the agent to reveal internal steps across unrelated tasks, including sensitive or security-relevant contexts. Because this is a data-analysis skill, users may also provide internal business data, making over-disclosure more risky.
This markdown file is written entirely in Chinese and presents the skill's examples and operating guidance as if Chinese is the required interaction language. The file does not indicate that users may choose another language or that the locale restriction is intentional and justified, which can violate language/locale choice policy.
This markdown file is entirely written as prescriptive guidance in Chinese and presents the analysis styles as instructions without any indication that users may choose another language or locale. Under the policy rule for natural-language violations, a skill that effectively forces a specific language without opt-in should be flagged unless the locale constraint is clearly documented and justified.
The manifest description and the entire instruction set are written as Chinese-only operating guidance, with no indication that the skill should adapt to the user's preferred language or locale. Under the policy, forcing a specific language without opt-in is a natural-language policy violation unless clearly documented and justified.
The requirement to '记下来' the user's correction and use it in later analysis is a plain-language instruction to store user-provided information for future reuse. While framed as helpful conversational continuity, it establishes memory/retention behavior over user inputs that can capture sensitive information in other contexts.
These trigger examples use vague phrases like '有什么值得注意的', '有没有关系', '定价策略对不对', and '有没有猫腻' without consistently requiring structured data input or clear data-analysis context. In a routing system, this can cause the skill to activate on broad everyday requests, leading to misrouting, unnecessary file/data access attempts, or over-collection of user data beyond the user's actual intent.
All natural-language examples in this file are written in Chinese, which can imply a language-specific behavior or evaluation constraint. There is no accompanying note that this skill is intentionally Chinese-only, region-specific, or that users may choose their preferred language.
No suspicious patterns detected.
