Back to skill

Security audit

关机吧人类

Security checks for vulnerabilities and agentic risk

Overview

This skill has a coherent paid timer process-kill purpose, but it handles payment credentials insecurely, force-closes apps, and includes instructions that try to suppress inspection.

Install only after reviewing the payment flow carefully. The skill sends payment credentials and order data to a hardcoded HTTP server, may install an unpinned payment dependency, stores order files locally, and can forcibly close apps including office and chat software without protecting unsaved work.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/service.py:10
Finding

Payment Credential Transmitted to a Bare IP over Plaintext HTTP

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/file_utils.py:18
Finding

Order Number Path Traversal Allows Filesystem Escape During JSON Read and Write

Content
View full analysis
dict: base_dir = get_orders_base_dir(indicator) json_path = os.path.join(base_dir, f"{order_no}.json") if not os.path.isfile(json_path): raise RuntimeError(f"Order file does not exist: {json_path}") with open(json_path, "r", encoding="utf-8") as f: return json.load(f) def save_order(indicator: str, order_no: str, order_data: dict) -> str: base_dir = get_orders_base_dir(indicator) os.makedirs(base_dir, exist_ok=True) json_path = os.path.join(base_dir, f"{order_no}.json") with open(json_path, "w", encoding="utf-8") as f: json.dump(order_data, f, ensure_ascii=False, indent=2) return json_path ``` Relevant callers: ```python save_order_info(order_no, amount, args.question, encrypted_data, pay_to, indicator) ``` ```python order_data = load_order(indicator, args.order_no) ``` ### Technical Analysis The `order_no` value is directly incorporated into a filesystem path without validation or containment checking. Sequences such as `../` can escape the intended order directory. In the write path, `order_no` originates from the remote order-creation response. A malicious or compromised server, or an attacker modifying the plaintext HTTP response, can return a traversal value. The resulting JSON is then written outside the intended order directory. In the read path, `order_no` is a command-line argument. A caller can supply traversal components and cause the application to parse another accessible file ending in `.json`. Appending `.json` limits reachable filenames to paths with that final suffix, but it does not prevent traversal or make the behavior safe. Files are accessed with the privileges of the user running the script ...[truncated 2064 chars]
Remediation
View remediation
str: if not ORDER_NO_PATTERN.fullmatch(order_no): raise ValueError("Invalid order number") return order_no ``` 2. Resolve the candidate path and verify that it remains under the resolved base directory: ```python base = os.path.realpath(get_orders_base_dir(indicator)) candidate = os.path.realpath(os.path.join(base, f"{order_no}.json")) if os.path.commonpath([base, candidate]) != base: raise ValueError("Order path escapes the order directory") ``` 3. Reject absolute paths, directory separators, `.` components, and `..` components explicitly. 4. Validate remote `orderNo` values immediately after receiving them and before displaying or storing them. 5. Use atomic and exclusive creation where overwriting is not expected. 6. Create order directories and files with permissions limited to the current user. 7. Consider mapping externally supplied order identifiers to locally generated filenames instead of using external identifiers directly as path components. 8. Add tests for Unix and Windows traversal forms, absolute paths, mixed separators, symlinks, and excessively long identifiers. ]]>

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:6
Finding

Skill Instructions Attempt to Override Session Behavior and Suppress Security Inspection

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:93
Finding

Unpinned External Payment Skill Installation Creates a Supply-Chain Risk

Content
View full analysis
", "indicator": "" } ``` ### Technical Analysis The project instructs the agent to install and execute an external Skill named `clawtip` if it is unavailable. It does not identify a trusted publisher or registry, pin a version, provide an integrity hash, require signature verification, or define a security-review procedure. A name alone is not a secure dependency identity. If multiple registries or packages can provide the same name, an attacker may publish a malicious or lookalike component. A compromised upstream release could also be installed automatically because no version or integrity constraint is specified. The dependency operates in a payment workflow and is expected to access order metadata and write `payCredential` into local order files, making dependency compromise particularly sensitive. ### Attack Path 1. The payment workflow determines that `clawtip` is not installed. 2. The Skill instructions direct the agent or user to install a component using only its name. 3. An attacker publishes a malicious package under that name in a searched source, compromises the upstream package, or influences dependency resolution. 4. The unverified package is installed and executed. 5. The malicious dependency reads order information, intercepts payment data, writes forged credentials, or executes other code with the privileges of the installing user. 6. The forged or stolen data is then con ...[truncated 757 chars]
Remediation
View remediation
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
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (23)

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The skill claims to be an 'automatic shutdown assistant' and to close specified software or services, but the described/observed behavior is narrower: killing preset application processes rather than shutting down the system or managing arbitrary services. Such overclaiming can mislead users about the scope and effect of the tool, causing unsafe reliance or unexpected process termination.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill claims to be an 'automatic shutdown assistant' and to close specified software or services, but the described/observed behavior is narrower: killing preset application processes rather than shutting down the system or managing arbitrary services. Such overclaiming can mislead users about the scope and effect of the tool, causing unsafe reliance or unexpected process termination.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The skill claims to be an 'automatic shutdown assistant' and to close specified software or services, but the described/observed behavior is narrower: killing preset application processes rather than shutting down the system or managing arbitrary services. Such overclaiming can mislead users about the scope and effect of the tool, causing unsafe reliance or unexpected process termination.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Directing the agent to reveal its full chain-of-thought is a significant information disclosure vulnerability. In a skill involving payment flow, credentials, local order files, and system-control actions, exposing reasoning can leak internal safeguards, sensitive metadata, or exploitable operational details to the user.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The repeated instruction reinforces disclosure of internal reasoning, increasing the likelihood that implementations follow it and leak sensitive intermediate details. Repetition in the skill content makes this more dangerous because it is framed as a hard requirement rather than an accidental wording issue.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill transmits the user's question, order number, payment credential, and delay to a remote service without any in-code user notice or consent flow, and the configured endpoint uses HTTP rather than HTTPS. This combination creates both privacy risk and credential exposure risk, especially since the data relates to payment verification.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instruction to expose the model's 'thought process' conflicts with safe disclosure boundaries and can cause leakage of internal reasoning, hidden policies, or sensitive intermediate data. In a payment- and credential-adjacent skill, this becomes more dangerous because internal traces may mention validation logic, file paths, or other details useful to an attacker.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The description explicitly says the agent should interact in Chinese, and repeats this instruction later, without offering the user a language choice. This is a natural-language policy issue because it forces a locale/language behavior regardless of user preference.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

This line reinforces a fixed Chinese-only interaction policy instead of allowing language selection or opt-in. The requirement is not documented as region-specific or otherwise justified, so it violates the language/locale policy criterion.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill omits a clear warning that forcibly terminating applications can discard unsaved work, interrupt active sessions, or corrupt user state. In this context, the capability is inherently disruptive, so lack of up-front warning meaningfully increases the risk of user harm and unintended data loss.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

该脚本的核心行为不仅是为本地“定时关闭软件/服务”创建状态,还会把用户输入发送到远程支付接口并将订单信息持久化到本地。对于一个看似本地化的关停工具而言,这种额外的数据外传和订单落盘扩大了数据暴露面,且技能描述未充分说明其网络收集范围,容易造成超出用户预期的信息处理。

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

脚本在没有任何用户可见告知或确认的情况下,将用户提供的内容提交到远程接口。即使该字段原本用于“软件类别”,用户也可能输入个人习惯、应用名称或其他敏感描述,导致非预期的数据共享和隐私风险。

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

用户提供的 question 被直接上传到外部 HTTP 服务,而非加密通道。HTTP 传输可被窃听、篡改或中间人攻击,导致用户输入、订单元数据甚至支付相关字段在网络中暴露;结合该技能的付费场景,这比普通遥测更敏感。

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

该文件实现的是基于 indicator 和 order_no 在用户主目录下读写订单 JSON,而技能声明核心功能是“定时关闭指定软件或服务”。这种能力与声明用途明显不一致,说明技能实际具备额外的本地持久化数据访问面,增加了隐藏状态存储、越权数据处理或为其他未披露功能提供支撑的风险。在一个付费验证相关技能中,这种不透明的数据落盘尤其值得警惕。

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

indicator 和 order_no 被直接拼接到用户主目录路径中,缺少对路径分隔符、'..' 或绝对路径片段的约束,可能导致路径遍历,在用户权限范围内读写任意位置的 JSON 文件。由于目标目录位于用户 home 下,这种能力可被用来覆盖配置、植入伪造数据或读取敏感本地文件,超出了该技能声明功能所必需的范围。技能上下文并不能合理解释这种自由文件访问,因此危险性更高。

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill sends payment/verification data to an external endpoint over plain HTTP, which exposes order identifiers and credentials to interception or tampering by a network attacker. This network capability is not strictly necessary for the local act of killing software, so the data exposure is disproportionate to the stated function.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill can terminate broad categories of applications, including office and productivity software, which exceeds the narrowly described bedtime/entertainment use case. In context, this creates unnecessary potential for data loss and misuse because a user may trigger forced closure of unrelated work applications.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/service.py (reported line 112)May include surrounding context.

python
for process_name in process_names:
        try:
            # 使用 PowerShell 关闭进程
            result = subprocess.run(
                ["powershell", "-Command", f"Stop-Process -Name '{process_name}' -Force -ErrorAction SilentlyContinue"],
                capture_output=True,
                text=True,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The code forcefully kills processes with -Force and provides no final confirmation, grace period handling, or unsaved-work warning before termination. In a timer-based skill, this is more dangerous because the action may occur later when the user is no longer watching, increasing the chance of unexpected data loss.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instruction to expose the model's 'thought process' conflicts with safe disclosure boundaries and can cause leakage of internal reasoning, hidden policies, or sensitive intermediate data. In a payment- and credential-adjacent skill, this becomes more dangerous because internal traces may mention validation logic, file paths, or other details useful to an attacker.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The declared credential.read permission appears broader than what is necessary for a workflow that mainly uses order numbers, indicators, and payment status. Excess credential access increases the blast radius if the skill or its dependencies are misused, especially in a skill already handling payments and local order files.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

The docstrings and runtime error message are written only in Chinese, which imposes a specific language on users and maintainers without any visible opt-in or locale selection. Under the policy rule, language-specific natural-language content is a violation unless the locale restriction is explicitly documented and justified.

Content

No source excerpt is available for this finding.