other
- Location
SKILL.md:40- Finding
Fabrication of Unsupported Clinical Facts in Generated Medical Records
- Content
View full analysis
Vulnerability Details
File Locations:
SKILL.md:40-45技能配置.md:46-51示例.md:31-74
Vulnerability Type: Unsafe medical record fabrication
Risk Level: HighVulnerable Instructions and Examples
SKILL.md:40-45:text ## 病历生成要求 严格按照以下模板生成专业、规范的入院记录： - 所有必填字段无遗漏 - 未明确说明的信息按"正常"或临床默认值填充 - 医学术语规范，避免口语化 - 确保病历的完整性与专业性技能配置.md:46-51:text ## 默认值规则 - 体温未提及：36.5℃ - 脉搏未提及：72次/分 - 呼吸未提及：18次/分 - 血压未提及：120/80mmHg - 其他未提及信息：填写"正常"或临床默认值Relevant portions of
示例.md:31-74demonstrate the resulting unsupported data:text 姓名：张无及 性别：男性 年龄：35岁 婚姻状况：未婚 民族：汉族 出生地：北京市 职业：职员 入院时间：2026-03-14 20:16 记录时间：2026-03-14 20:16 病史陈述者：患者 (可靠程度：可靠) **既往史**：既往体健；否认"高血压、糖尿病、冠心病"等慢性病史。否认"乙肝、结核、艾滋、梅毒"等传染病史。无手术、外伤及输血史。否认食物、药物过敏史。疫苗接种情况不详。 **体格检查**： 生命体征：T: 38.5℃，P: 85次/分，R: 20次/分，BP: 120/80mmHg。 **初步诊断**： 1. 急性上呼吸道感染 2. 发热待查：感染性？Technical Analysis
The Skill explicitly instructs the Agent to replace missing clinical information with normal findings or predetermined values. This violates a fundamental medical-record integrity requirement: an unobserved or unreported finding must not be represented as measured, examined, or affirmatively denied.
The example confirms that this is not limited to harmless formatting defaults. From an input containing only fever, cough, duration, sex, and age, the generated record adds an identity, marital status, ethnicity, birthplace, occupation, timestamps, reliability assessment, detailed negative medical history, physical examination findings, vital signs, and preliminary diagnoses.
The generated diagnoses also conflict with
SKILL.md:48, which states that the Skill should only generate records and should not provide diagnostic advice. No provenance markers distinguish user-supplied facts from inferred or fabricated content.The issue does not grant operating-system privileges, execute code, or obtain unaut ...[truncated 1906 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove all instructions that replace missing clinical information with normal values or affirmative negative findings.
- Represent unavailable data explicitly as
Not provided,Unknown,Not assessed, orNot examined, as appropriate. - Require explicit user confirmation for every measurement, examination result, medical-history denial, demographic fact, and diagnosis.
- Preserve provenance by labeling each field as user-provided, clinician-observed, imported, or unresolved.
- Do not generate diagnoses unless separately requested within an appropriately controlled workflow. Any generated differential should be clearly labeled as unverified and must not be presented as an established diagnosis.
- Add validation that rejects generation when mandatory factual fields are absent rather than silently inventing them.
- Replace the unsafe example with one that leaves unspecified fields unresolved and clearly labels the document as an unverified draft requiring qualified clinician review.
- Add automated tests using sparse inputs to verify that the Skill never fabricates vital signs, examination results, history, identity data, timestamps, or diagnoses.