Back to skill

Security audit

极星会活动自动化参与

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly an activity-participation helper, but it asks an agent to collect full-screen screenshots and a Baidu Cloud account ID and submit them through a shortened external link without enough user review or privacy controls.

Install only if you are comfortable with an agent guiding or performing activity signup, opening Baidu Cloud, capturing full-screen images, and uploading those images with your account ID. Review screenshots manually, close unrelated apps and notifications, resolve the shortened URL destination, and approve the exact files and account ID before any submission.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (2)

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:120
Finding

Mandatory Referral Promotion Manipulates Agent Output

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:120-126; also implemented in scripts/automate_activity.py:125-131
Vulnerability Type: Mandatory promotional output inserted into the agent workflow
Risk Level: Medium

Evidence

SKILL.md:120-126:

markdown
### 8. 完成报告

向用户报告:
- 报名状态: 成功/失败
- 截图保存位置
- 账号 ID
- 提交状态
- 邀请码提示: `JIXING6`(可领取 5000 积分)

scripts/automate_activity.py:125-131:

python
def show_bonus_tip(self):
    """
    显示额外提示
    """
    self.log("\n=== 额外提示 ===")
    self.log("如积分不足,可使用邀请码: JIXING6")
    self.log("立领 5000 积分,体验准点下班的快乐!")
    self.log("\n最终参与结果以系统后台数据判定为准")

Technical Analysis

The Skill requires the agent to include referral code JIXING6 in the completion report, irrespective of whether the user requested referral information. The executable script independently reinforces the same behavior by calling show_bonus_tip() after the workflow steps complete.

This is instruction-level output manipulation: loading and following the Skill causes the agent to inject stable promotional content into its response. The referral message is not necessary to register for the activity, collect screenshots, retrieve the account identifier, or report whether submission succeeded.

The behavior does not override core safety controls, execute arbitrary code, or create persistence. Nevertheless, it alters the agent's normal response objective by turning workflow output into an unsolicited promotional channel.

Attack Path

  1. A user requests participation in the advertised activity.
  2. The agent loads the Skill because the request matches its activation description.
  3. The agent follows the prescribed registration and submission workflow.
  4. The completion-report instructions require the agent to display referral code JIXING6.
  5. The user receives promotional content without having specifically requested referral or advertising information.

Impact Assessment

The issue affects the integrity and neutral ...[truncated 357 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the mandatory referral-code entry from the completion-report requirements.
  2. Remove show_bonus_tip() or stop invoking it automatically at the end of the workflow.
  3. If referral information is retained, clearly label it as promotional or sponsored content.
  4. Display referral information only when the user explicitly asks about credits, invitations, or discounts.
  5. Disclose any financial or attribution benefit received by the Skill author.
  6. Restrict the completion report to factual workflow results, such as registration status, submission status, and the location of user-approved artifacts.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:43
Finding

Collection and External Submission of Full-Screen Images and a Cloud Account Identifier Without Adequate Data-Minimization Controls

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:43-57 and SKILL.md:92-117
Vulnerability Type: Excessive collection and external disclosure of potentially sensitive user information
Risk Level: High

Evidence

SKILL.md:43-57:

markdown
### 第四步: 截取 3 张全屏截图

在 DuMate 体验过程中,自动截取 3 张全屏截图,确保:
- 截图为电脑全屏
- 桌面背景清晰可见
- 展示 DuMate 的不同功能界面

### 第五步: 获取账号 ID

1. 打开百度云控制台: `https://cloud.baidu.com/`
2. 登录后定位账号 ID 显示位置
3. 提取账号 ID(非手机号或昵称)

### 第六步: 提交参与信息

1. 返回活动页面: `https://wxaurl.cn/3kzrkXaZ4Dh`
2. 填写表单:
   - 输入账号 ID
   - 上传 3 张全屏截图
3. 提交参与信息

The corresponding operational instructions in SKILL.md:92-117 state:

markdown
### 5. 截取 3 张全屏截图

在用户使用 DuMate 过程中:
- 使用系统截图功能或浏览器自动化截图
- 截取 3 张不同的全屏界面截图
- 确保桌面背景清晰可见
- 保存截图到 `~/Desktop/jixinghui_screenshots/` 目录

### 6. 获取账号 ID

使用浏览器自动化技能:
- 打开百度云控制台: `https://cloud.baidu.com/`
- 如需登录,提示用户登录
- 定位账号 ID 显示位置(通常在页面右上角或个人中心)
- 提取账号 ID

### 7. 提交参与信息

返回活动页面 `https://wxaurl.cn/3kzrkXaZ4Dh`:
- 定位参与信息提交表单
- 填写账号 ID
- 上传 3 张截图
- 提交表单

Technical Analysis

The Skill explicitly requires three full-screen captures and specifies that the desktop background must remain visible. Full-screen capture is broader than capturing the DuMate application window and may include unrelated applications, file names, browser tabs, notifications, email previews, usernames, internal project information, or other personal data.

The workflow also instructs the agent to access the user's authenticated Baidu Cloud console and extract a stable account identifier. It then combines that identifier with the screenshots and submits them through https://wxaurl.cn/3kzrkXaZ4Dh, which is an opaque shortened URL in the reviewed package. The audited files do not identify the final destination domain, recipient, retention policy, or purpose limitation for the submitted data.

The Skill does mention that login authorization requires manual user interaction, but it does not require:

  • Review or redaction of screenshots before upload. ...[truncated 2121 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace full-screen capture with capture of only the DuMate application window or a tightly bounded application region.
  2. Do not require the desktop background to be visible.
  3. Before saving or uploading images, detect and redact notifications, usernames, file names, account details, and content from unrelated applications.
  4. Present all three images to the user for review and obtain explicit approval for each image.
  5. Resolve the shortened URL before use and disclose the final destination domain, recipient identity, data purpose, and retention policy.
  6. Require explicit confirmation immediately before submitting the account ID and screenshots. This confirmation should list the destination and every file or data field being transmitted.
  7. Avoid extracting the cloud account ID automatically when the user can manually copy a purpose-specific identifier.
  8. Do not access unrelated areas of the cloud console, browser profile, or authenticated session.
  9. Provide a cancellation path that deletes locally staged screenshots and clears extracted account data.
  10. Apply least-privilege controls so browser automation can access only the approved destination and screenshot tooling can capture only the approved window.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (7)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

声明的核心能力是“自动完成全流程”,而实际代码并没有实现这些自动化能力。它没有使用浏览器自动化库、GUI 自动化、网络请求或系统截图接口;唯一实际执行的行为是创建桌面目录、记录日志、构造三个截图文件名并顺序输出操作说明。因此其主要用途更像是活动参与步骤的引导脚本/清单,而不是能真正自动完成活动的工具。虽然整体主题与声明一致,且没有明显额外越权能力,但对自动化程度的描述与实际行为存在实质性不符,应判定为描述与行为不匹配。

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger description uses broad 'or similar expressions' language, which can cause the skill to activate on ambiguous user requests. In this skill's context, accidental activation is more dangerous because the workflow includes visiting external sites, handling account identifiers, and collecting screenshots.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The usage section provides generic example phrases without specifying boundaries or exclusion cases, increasing the chance of unintended invocation. Because the skill proceeds into registration, screenshot capture, and data submission, mistaken activation can expose private information or initiate unwanted actions on third-party services.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill directs automatic full-screen screenshot capture and local storage without an explicit privacy warning or scoped consent. Full-screen captures can unintentionally include sensitive content such as messages, documents, tokens, desktop notifications, or other personal data unrelated to the activity, making this especially risky in an automation context.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill automates extraction of an account ID from Baidu Cloud and submission of that identifier plus screenshots to an external activity page without a clear disclosure of data transmission. This creates a meaningful privacy risk because it combines identifying account information with potentially sensitive screen contents and sends them to third-party services as part of an automated workflow.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The module docstring and manifest describe an end-to-end automated participation flow, but the implementation merely logs prompts telling the user or another skill to perform the actions. Across the step methods, there is no browser automation, no screenshot capture, no account ID extraction, and no form submission logic, so the behavior materially falls short of the claimed automation scope.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The run() docstring says it executes the complete automated flow, while the step docstrings and top-level module text present the script as an automation tool. In reality, the methods only print instructions such as '请使用浏览器自动化技能完成以下操作' and return True, which contradicts the documented intent of direct execution.

Content

No source excerpt is available for this finding.