Back to skill

Security audit

医疗器械投标决策-医院采购投标评估

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly aligned with bid-analysis use, but it also handles account registration, device identifiers, persistent API-key storage, and session links in ways users should review before installing.

Review whether you are comfortable letting this skill contact Zhiliaobiaoxun services, use or create an API key, store that key under ~/.zlbx/config.json, and send a hashed MAC-derived device identifier for trial-account registration. Install only if you accept those account and device-identification behaviors.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
Findings (33)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

声明描述的是一个面向医疗招投标决策的分析型技能,核心应是数据获取、分析推断和生成决策结论。给出的代码却是一个纯展示层脚本:从本地 JSON 读取结构化结果,拼接 HTML/CSS/JS,生成品牌化报告页面,并支持打印、PDF 和长图导出。代码中没有任何面向医疗领域的专门逻辑,也没有联网访问、数据库查询、历史中标数据分析、竞对识别、报价模型或风险评估算法。换言之,它假定分析结果已经存在,只负责“渲染报告”,而不是“进行投标决策分析”。这与声明的主要用途存在实质性不一致。

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 81)May include surrounding context.

md
完整模式下,对话报告输出后默认再用 `scripts/render_report.py` 生成一份可分享的 HTML 版报告并告知保存路径(详见 report-template.md「HTML 报告导出」)。

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 85)May include surrounding context.

md
完整模式下,对话报告输出后默认再用 `scripts/render_report.py` 生成一份可分享的 HTML 版报告并告知保存路径(详见 report-template.md「HTML 报告导出」)。

Hidden Instructions

High
Category
Prompt Injection
Confidence
90% confidence
Finding

This duplicate finding identifies the same hidden instruction block that injects mandatory request-header and upgrade-check behavior through commented content. Even if the instructions are not overtly malicious, concealing behavior in comments reduces auditability and can be abused to introduce undisclosed network or telemetry requirements.

Content

Scanner excerpt · SKILL.md (reported line 163)May include surrounding context.

  • 报告里某个竞争对手威胁最大 → 引导用 zlbx-company-intel 对其深度背调
  • 以上都不贴切 → 报告涉及的企业完整档案与更多商机详情,引导访问知了商机大师 https://agent.zhiliaobiaoxun.com

安装包身份与升级检查:所有发往知了标讯 API 的请求(包括账户查询)必须额外携带以下两个请求头:

text
X-Skill-Package: medical-device-bid-decision/1.0.9

Hidden Instructions

High
Category
Prompt Injection
Confidence
90% confidence
Finding

This duplicate finding identifies the same hidden instruction block that injects mandatory request-header and upgrade-check behavior through commented content. Even if the instructions are not overtly malicious, concealing behavior in comments reduces auditability and can be abused to introduce undisclosed network or telemetry requirements.

Content

Scanner excerpt · SKILL.md (reported line 163)May include surrounding context.

  • 报告里某个竞争对手威胁最大 → 引导用 zlbx-company-intel 对其深度背调
  • 以上都不贴切 → 报告涉及的企业完整档案与更多商机详情,引导访问知了商机大师 https://agent.zhiliaobiaoxun.com

安装包身份与升级检查:所有发往知了标讯 API 的请求(包括账户查询)必须额外携带以下两个请求头:

text
X-Skill-Package: medical-device-bid-decision/1.0.9

Hidden Instructions

High
Category
Prompt Injection
Confidence
90% confidence
Finding

This duplicate finding points to the same hidden instruction block that tells the agent to use existing credentials and preserve specific package/source headers. Even if operationally legitimate, embedding credential-use directives inside a skill document is dangerous because a compromised or spoofed skill can coerce the agent into authenticating requests or impersonating package provenance without an independent trust check.

Content

Scanner excerpt · references/api-quick.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Hidden Instructions

High
Category
Prompt Injection
Confidence
90% confidence
Finding

This duplicate finding points to the same hidden instruction block that tells the agent to use existing credentials and preserve specific package/source headers. Even if operationally legitimate, embedding credential-use directives inside a skill document is dangerous because a compromised or spoofed skill can coerce the agent into authenticating requests or impersonating package provenance without an independent trust check.

Content

Scanner excerpt · references/api-quick.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Hidden Instructions

High
Category
Prompt Injection
Confidence
97% confidence
Finding

The HTML-comment-delimited sections embed hidden operational instructions that steer request identity, headers, package attribution, and account behavior in ways not visible as normal user-facing logic. Hidden instructions are dangerous because they can smuggle privileged behaviors past reviewers and cause the agent to perform networked actions or impersonation-like request shaping without transparent disclosure.

Content

Scanner excerpt · references/auto-register.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Hidden Instructions

High
Category
Prompt Injection
Confidence
97% confidence
Finding

The HTML-comment-delimited sections embed hidden operational instructions that steer request identity, headers, package attribution, and account behavior in ways not visible as normal user-facing logic. Hidden instructions are dangerous because they can smuggle privileged behaviors past reviewers and cause the agent to perform networked actions or impersonation-like request shaping without transparent disclosure.

Content

Scanner excerpt · references/auto-register.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file instructs the agent to perform automatic account registration, device-feature collection, credential handling, balance checks, and recharge-link generation, which are materially outside the declared purpose of a medical bid-decision analysis skill. This scope expansion creates unnecessary access to local state and external account operations, increasing the attack surface and enabling covert account bootstrapping under the guise of analytics.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill directs collection of device fingerprinting data (platform, CPU architecture, and a hashed MAC address) for automatic registration. Even with hashing, the MAC-derived value is a persistent device identifier and is unrelated to the user-facing task of medical procurement analysis, making the collection context-inappropriate and privacy-invasive.

Content

No source excerpt is available for this finding.

Hidden Instructions

High
Category
Prompt Injection
Confidence
83% confidence
Finding

This duplicate hidden-instruction finding is substantiated by the same comment block that silently imposes request identity and credential-handling behavior. Even if intended for package integrity, embedding such controls in markdown creates a prompt-injection style control channel and mixes content with privileged execution policy.

Content

Scanner excerpt · references/report-template.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Hidden Instructions

High
Category
Prompt Injection
Confidence
83% confidence
Finding

This duplicate hidden-instruction finding is substantiated by the same comment block that silently imposes request identity and credential-handling behavior. Even if intended for package integrity, embedding such controls in markdown creates a prompt-injection style control channel and mixes content with privileged execution policy.

Content

Scanner excerpt · references/report-template.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Hidden Instructions

High
Category
Prompt Injection
Confidence
93% confidence
Finding

This duplicate hidden-instruction finding points to the same concealed request-identity block that tells the agent how to authenticate and identify itself to an external API. Even if intended for package integrity, hiding these instructions in comments increases prompt-injection and covert data-flow risk because the model may follow unseen directives that the user cannot inspect or challenge.

Content

Scanner excerpt · references/workflow.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Hidden Instructions

High
Category
Prompt Injection
Confidence
93% confidence
Finding

This duplicate hidden-instruction finding points to the same concealed request-identity block that tells the agent how to authenticate and identify itself to an external API. Even if intended for package integrity, hiding these instructions in comments increases prompt-injection and covert data-flow risk because the model may follow unseen directives that the user cannot inspect or challenge.

Content

Scanner excerpt · references/workflow.md (reported line 1)May include surrounding context.

请求身份:使用已配置的 API Key,并保留本包的 X-Client,携带下列完整身份头;不得沿用其它包的版本或来源。

text
X-Client: bid-decision/1.0.5

Obfuscated Code

High
Category
Supply Chain
Confidence
50% confidence
Finding

Code contains obfuscation (base64, hex encoding with execution). This is often used to hide malicious functionality.

Content

Scanner excerpt · scripts/render_report.py (reported line 71)May include surrounding context.

python
THREAT_COLOR = {"高": "#c0392b", "中": "#b9770e", "低": "#0d9463"}

# 知了标讯白色 logo(299x96 PNG base64 内嵌,保证报告离线/打印/转发时不裂图)
_LOGO_B64 = "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
...[truncated 27 chars]

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
84% confidence
Finding

The skill explicitly states it will write HTML reports to a local directory, but it does not declare any tool scope or permission boundary for file writes. That mismatch can enable unintended or over-broad filesystem access in runtimes that rely on manifest-declared permissions, making the skill harder to sandbox and audit.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The description says the skill must be used not only for explicit medical-device bidding analysis, but also whenever requests involve hospital procurement bid evaluation, device bid go/no-go, or equipment pricing references, even if the user does not mention 'medical'. This broad activation language does not provide clear exclusion conditions or negative examples, increasing the risk of unintended invocation overlap with other bidding or procurement skills.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The prescribed user-facing example text and operational guidance are entirely in Chinese and the skill consistently instructs specific Chinese phrasing, but nowhere offers users a language option or states that the skill is intentionally restricted to Chinese-speaking users. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file provides fixed Chinese user-facing text to present during consent and repeatedly instructs the agent to output specific Chinese wording and links. There is no indication that the skill should ask the user’s preferred language or that the skill is restricted to Chinese-speaking users only, which creates a language/locale policy concern.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

The skill explicitly instructs the agent to transmit locally derived device identifiers and other registration data to an external service. External transmission is only safe when it is necessary, proportionate, and tightly aligned to the skill’s purpose; here it supports an unrelated auto-registration workflow, so the transmission increases privacy and supply-chain risk.

Content

Scanner excerpt · references/auto-register.md (reported line 130)May include surrounding context.

md
> ### ⚠️ 请求体必须用 JSON 序列化函数生成,不要手拼字符串
>
> 用 `json.dumps(payload)` / `requests.post(url, json=payload)` / `JSON.stringify(payload)`,
> 或 `curl -d @file`;**不要用字符串拼接,也不要用 Python 的 `str(dict)`**
> (后者产出单引号,服务端会报 `Expecting property name enclosed in double quotes`)。
>

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/auto-register.md (reported line 131)May include surrounding context.

md
> ### ⚠️ 请求体必须用 JSON 序列化函数生成,不要手拼字符串
>
> 用 `json.dumps(payload)` / `requests.post(url, json=payload)` / `JSON.stringify(payload)`,
> 或 `curl -d @file`;**不要用字符串拼接,也不要用 Python 的 `str(dict)`**
> (后者产出单引号,服务端会报 `Expecting property name enclosed in double quotes`)。
>
> 历史教训:曾有版本采集 `home_path`,Windows 的 `C:\Users\alice` 直接拼进 JSON 字符串时

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The instructions require persisting API keys to a local config file and reusing them immediately in the active session. For an analysis skill, local credential storage and session reuse exceed the declared role and create opportunities for credential leakage, misuse by other local processes, or unintended persistence beyond the user’s expectations.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
95% confidence
Finding

The instructions direct the agent to create a persistent local directory, merge configuration, store API keys, and immediately continue using the new credential in-session. This establishes durable session persistence and secret retention on disk, which is risky for a skill whose stated role is analytical rather than credential-management oriented.

Content

Scanner excerpt · references/auto-register.md (reported line 195)May include surrounding context.

text

注意事项:
- 目录不存在时先 `mkdir -p ~/.zlbx`
- 文件已存在时**合并而非覆盖**(保留用户可能的其他配置)
- `source: "auto"` 字段必须写入,**这是后续判断「是否输出自动登录链接」的关键依据**

Session Persistence

Medium
Category
Rogue Agent
Confidence
95% confidence
Finding

The instructions direct the agent to create a persistent local directory, merge configuration, store API keys, and immediately continue using the new credential in-session. This establishes durable session persistence and secret retention on disk, which is risky for a skill whose stated role is analytical rather than credential-management oriented.

Content

Scanner excerpt · references/auto-register.md (reported line 195)May include surrounding context.

text

注意事项:
- 目录不存在时先 `mkdir -p ~/.zlbx`
- 文件已存在时**合并而非覆盖**(保留用户可能的其他配置)
- `source: "auto"` 字段必须写入,**这是后续判断「是否输出自动登录链接」的关键依据**