T05 · Unauthorized Access and Privilege Escalation
- Location
SKILL.md:15- Finding
Planning Document Access Violates the Declared Isolation Boundary
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is not malware, but it should be reviewed because it handles sensitive child and family feedback with under-scoped document access and forwarding rules.
Install only if this workflow is meant for a controlled Chinese-speaking parent/planner process, parents have been told their replies may be forwarded to the planner, and the planning document metadata boundary is fixed or explicitly accepted. Prefer moving plan_meta to a separate state store, requiring plan_id confirmation before sends or reply processing, and minimizing or redacting sensitive feedback before forwarding when possible.
SKILL.md:15Planning Document Access Violates the Declared Isolation Boundary
The skill claims it does not read or modify planning report documents, but the workflow explicitly instructs it to read plan_meta from the planning document and later update that metadata. This creates a trust-boundary violation: an operator or user may rely on the stated limitation, while the implemented workflow still accesses and mutates report-adjacent data, increasing the chance of unauthorized document access or accidental report corruption.
This is a direct contradiction between the core rule ('不碰报告') and the operational steps that read and update plan_meta in the planning document. Such inconsistencies are security-relevant because they can bypass user/operator expectations, undermine review controls, and lead to unintended handling of sensitive child planning data.
The trigger list includes broad natural phrases such as '复盘一下' and '跟进一下孩子', which could be used in ordinary conversation and unintentionally invoke the skill. In a child-related workflow that sends questionnaires and forwards parent feedback, accidental activation can expose sensitive family information or trigger messages without clear consent.
The skill description, prompts, and user-facing messages are entirely in Chinese, and no opt-in or alternative language behavior is offered. Under the stated policy, forcing a specific language without user choice or explicit documented justification is a natural-language policy violation.
The skill description and flow do not clearly warn parents that their feedback, including potentially sensitive or high-risk content, may be forwarded to the planner and, for red-line cases, sent immediately with original text. This is a transparency and consent failure involving sensitive child/family data, which can create privacy harm and violate user expectations.
The reply-processing trigger of '收到家长回复 / 手动说"处理跟进回复"' is too broad and may capture unrelated parent messages or operator requests without sufficient scoping. Because this mode reads raw parent replies, performs high-risk screening, and forwards content to a planner, ambiguous invocation raises the risk of misrouting private communications.
The file consists entirely of end-user questionnaire and reminder templates written in Chinese, including headings and user-facing messages, with no indication that another language can be used or that the user has opted into Chinese communication. This creates a natural-language locale policy issue under the rule for skills that force a specific language without user choice.
The confirmation template says the feedback has been forwarded to the planner and that they will evaluate it and provide updated advice, which directly contradicts the skill manifest stating it does not perform evaluation or strategy adjustment. This can cause scope creep, mislead users about the service being provided, and trigger downstream human or agent behavior outside approved boundaries.
This documentation explicitly states that the feedback will be evaluated and used to update advice, contradicting the declared no-evaluation/no-strategy-adjustment scope. In a child development workflow, such contradictory instructions can lead operators or automation to provide unapproved guidance, creating compliance, trust, and safety risks.
The red-line procedure requires immediate forwarding of the parent's full original message, which may include highly sensitive family, behavioral, medical, and safety information, but the file provides no privacy notice, minimization rule, or disclosure to the user about sharing. In a child-development context, this increases the risk of over-collection and unnecessary exposure of sensitive child and family data during escalation.
The forwarding templates package child name, exact age, plan_id, follow-up count, and potentially sensitive parent feedback, including full unredacted text in red-line cases, without any accompanying privacy guidance or consent language. Because the skill serves children aged 0–8 and handles medical/family-risk signals, the omission materially raises privacy and data-sharing risk.
The template tells the planner to '评估后更新规划文档', which operationally extends this follow-up skill into a workflow that triggers evaluation and plan-document updates despite the manifest explicitly saying it does not do evaluation or modify reports. That scope mismatch is dangerous because it can cause downstream automation or humans to rely on this skill for actions that require stricter authorization, review, and consent boundaries.
All prescribed user-facing responses and templates are fixed in Chinese, and the document does not offer a language choice or note that the skill is intentionally limited to Chinese-speaking users. This is a natural-language policy concern because it forces a specific language without documented user opt-in.