Back to skill

Security audit

推送消息数据助手(友盟U-Push)

Security checks for vulnerabilities and agentic risk

Overview

This skill is presented as read-only, but it handles powerful session cookies in unsafe ways and includes overbroad authenticated request capabilities.

Review carefully before installing. Avoid pasting full session Cookies into chat, do not run the automatic Cookie extraction or custom request commands, and treat any saved Cookie as equivalent to account access. If used anyway, restrict it to a disposable/least-privilege Umeng account and rotate Cookies or vendor secrets that may have appeared in logs.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Tool Hijacking and SpoofingModifies or replaces tools so legitimate-looking calls execute attacker logic
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (6)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/api_request.py:42
Finding

Authenticated Session Cookie Can Be Sent to an Arbitrary URL

Content
View full analysis
[method]") sys.exit(1) url = sys.argv[2] method = sys.argv[3] if len(sys.argv) > 3 else 'GET' result = custom_request(url, method.upper()) ``` ### Technical Analysis The `custom` command accepts a user-controlled URL and HTTP method. Before making the request, `make_request()` loads the complete saved Umeng session Cookie and places it in the `Cookie` header. It also includes the CSRF token extracted from that Cookie. There is no validation of the URL scheme, hostname, port, path, resolved IP address, or redirect destination. Consequently, the authentication headers are not limited to `https://upush.umeng.com`. The separate `security_interceptor.py` ...[truncated 1554 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
scripts/security_interceptor.py:58
Finding

Documented Read-Only API Restrictions Are Unenforced and Fail Open

Content
View full analysis
Remediation
View remediation

T07 · Tool Hijacking and Spoofing

Error
Location
scripts/auto_get_cookie.py:30
Finding

Obsolete Script Automatically Extracts Browser Authentication Cookies Through MCP

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/manage_cookie.py:11
Finding

Session Cookie Is Stored Without Enforced Permissions and Can Be Printed in Full

Content
View full analysis
2: ctoken = extract_ctoken(sys.argv[2]) if ctoken: print(ctoken) else: print("NO_CTOKEN", file=sys.stderr) sys.exit(1) ``` ### Technical Analysis The primary Cookie manager writes a reusable session credential as plaintext without explicitly applying owner-only permissions. The final mode therefore depends on the process umask and preexisting file permissions. This is inconsistent with `scripts/save_cookie.py`, which applies mode `0600`. The `load` command prints the complete Cookie to standard output, while `extract-ctoken` prints the CSRF token. Standard output is commonly captured by Agent transcripts, terminal logging, CI systems, shell wrappers, or support logs. These commands therefore create direct credential-disclosure paths. The project also maintains two different credential locations—`scripts/cookie.json` and `~/.qoderwork/skills/umeng-push-helper/cookie.txt`—which can leave stale credential cop ...[truncated 920 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/query_daily_api_stats.py:460
Finding

Daily Statistics Command Prints Vendor Push Credentials

Content
View full analysis
1 else fields[0], '-' ) or '-' if len(fields) == 3: app_id_val = field_values.get(fields[0], '-') or '-' secret_val = field_values.get(fields[2], '-') or '-' callback_val = field_values.get(fields[1], '-') or '-' if len(fields) == 1: app_id_val = '-' secret_val = field_values.get(fields[0], '-') or '-' callback_val = '-' def truncate(val, max_len=33): val = str(val) if val else '-' if len(val) > max_len: return val[:max_len] + '..' return val print( f"{vendor['name']:<10} " f"{truncate(app_id_val):<35} " f"{truncate(secret_val):<35} " f ...[truncated 1444 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/query_msg_detail.py:586
Finding

Generated HTML Report Allows Script and HTML Injection from API-Controlled Data

Content
View full analysis
单条消息详情 - {msg_id} ``` ```javascript const data = {data_json}; const infoFields = [ ['任务描述', data.msgInfo.description || data.msgInfo.title || ''], ['推送类型', data.msgInfo.type || data.msgInfo.pushType || ''], ['目标人群', data.msgInfo.target || ''], ['生产模式', data.msgInfo.productionMode ? '是' : '否'], ['发送时间', data.msgInfo.pushTime || data.msgInfo.sendTime || ''], ['创建时间', data.msgInfo.createTime || ''], ['状态', data.msgInfo.status || ''], ['厂商通道配置', data.msgInfo.channelActivity ? '已配置' : '未配置'], ]; infoFields.forEach(([label, value]) => { if (value) { const div = document.createElement('div'); div.className = 'info-item'; div.innerHTML = `
${label}
${value}
`; infoGrid.appendChild(div); } }); ``` ```javascript html += ` ${item.stage || item.phase || ''} ${getChName(item.channel || '')} ${item.reason || item.cause || item.errorMsg || ''} ${count.toLocaleString()} ${ratio} `; ``` ```python with open(filepath, 'w', enc ...[truncated 2221 chars]
Remediation
View remediation
`, `&`, Unicode line separators, and ``, or place the data in a non-executable JSON block and parse it safely. 4. Escape `msg_id` before placing it in the HTML title or any markup context. 5. Add a restrictive Content Security Policy, for example disallowing inline scripts and outbound connections where feasible. 6. Bundle a reviewed, pinned Chart.js asset locally rather than loading executable code from a CDN. 7. If a CDN remains necessary, use Subresource Integrity and a restrictive `crossorigin` policy. 8. Add regression tests using payloads containing ``, quotes, backticks, ampersands, and ``. ]]>
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (99)

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The changelog documents a shift from user-supplied credentials to the skill automatically opening a browser, reading authenticated cookies, and persisting them locally. That materially expands the skill's authority beyond its declared read-only helper scope and creates a credential-harvesting behavior that could expose session tokens if the local file is accessed or reused improperly.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Automatically opening a browser and executing JavaScript to read browser cookies is a sensitive capability unrelated to simple read-only push-query functionality. In this context, the skill is handling live authentication material from a user's browser session, which increases the risk of credential theft, session hijacking, and misuse far beyond the stated business purpose.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill is described as a read-only Umeng query helper, but the examples explicitly show authenticated POST requests to a send endpoint and a generic custom request capability. This creates a dangerous mismatch between the declared trust boundary and actual capability, making it easy for an operator or downstream agent to perform state-changing actions under the user's authenticated session.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and automatically opening local HTML reports is not inherently dangerous, but it is behavior outside the narrowly described query role and it adds local execution surface through OS-specific open commands. In a skill already handling sensitive authenticated data, extra local side effects reduce predictability and could expose report contents to unintended viewers on a shared machine.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill tells users not to share cookies in chat or logs, yet instructs them to paste the full authentication cookie directly into the conversation. That contradiction creates a direct credential-exfiltration path through chat history, telemetry, model context, screenshots, or support logs.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Instructing users to paste full session cookies into chat exposes highly sensitive bearer credentials in a medium not designed for secret handling. Anyone with access to conversation records or model traces could reuse the cookie to impersonate the user in the Umeng backend.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill creates a natural-language secret leakage path by asking for full authentication cookies in chat and then saving them locally. Because session cookies are bearer credentials, disclosure enables account takeover or unauthorized backend access until expiry or revocation.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document gives step-by-step instructions for revealing and copying full browser session cookies from developer tools without a strong up-front warning that these credentials can grant account access if stolen. In practice, this normalizes unsafe handling of bearer tokens and increases the likelihood of compromise by phishing, shoulder-surfing, clipboard leakage, or transcript capture.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Supporting commands that load and show the complete saved cookie value re-exposes a sensitive secret after storage and increases the chance it will appear in terminal history, logs, screen recordings, or copied output. This turns local persistence into repeated credential disclosure.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The error-handling workflow reinforces repeated submission of fresh full cookies through chat whenever authentication fails. Repeated collection compounds exposure, increases the number of credential copies in transcripts, and trains users into unsafe secret-sharing behavior.

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 750)May include surrounding context.

md
- `SKILL.md` - 技能定义文档

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 763)May include surrounding context.

md
- `scripts/close_attribution_template.html` - 关闭归因 HTML 图表模板

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 764)May include surrounding context.

md
- `scripts/overview_stats_template.html` - 概况统计 HTML 图表模板

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill is described as a read-only Umeng query helper, but the script explicitly exposes non-read-only functionality such as send_test in the usage text and a generic request path that can issue POST requests. In a credential-bearing tool that automatically attaches saved session cookies, this creates a privilege and trust-boundary violation: users may believe they are only querying data while the tool can perform state-changing actions against the Umeng backend.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The custom command accepts an arbitrary URL and passes it to make_request, which automatically includes the saved Umeng authentication cookie and CSRF token in headers. This turns the tool into a credential forwarding primitive that can send valid session material to attacker-controlled domains or unintended third parties, far beyond the stated read-only helper purpose.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
100% confidence
Finding

custom_request allows a caller to specify any URL, and make_request will send the persisted Cookie header and derived CSRF token to that destination without warning or origin validation. Because the skill requires users to supply browser session cookies, this is especially dangerous: it can directly exfiltrate authentication material to an attacker-controlled server and enable account compromise.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The code uses browser navigation plus JavaScript evaluation specifically to retrieve authentication cookies, which is a powerful credential-access capability unrelated to the stated read-only business purpose. Even if intended for convenience, this creates a high-risk path for session theft, impersonation, and unauthorized access if the cookie is exposed, reused, or mishandled.

Content

No source excerpt is available for this finding.

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · scripts/close_attribution_template.html (reported line 147)May include surrounding context.

html
</div>
        </div>
        
        <!-- 模块 1: 用户偏好分析 -->
        <div class="section">
            <div class="section-title">👤 用户偏好分析</div>
            <div class="block">

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The code path is capable of accessing sensitive vendor configuration data and later rendering it for the operator, which exceeds the stated purpose of querying push statistics. Because the skill requires a live authenticated Umeng Cookie, this behavior operates with real account privileges and can disclose third-party vendor secrets that may allow unauthorized API use or account takeover in connected push providers.

Content

No source excerpt is available for this finding.