Back to skill

Security audit

管理无阶未来GPU云平台实例的创建、查询与删除操作;当用户需要创建GPU容器、查询实例状态或清理资源时使用

Security checks for vulnerabilities and agentic risk

Overview

The skill matches its stated GPU cloud management purpose, but it can delete authenticated cloud instances immediately without a built-in confirmation or preview step.

Review this before installing if an agent may act autonomously. Only use it with a Vylai token whose permissions and account balance you are comfortable exposing, and require explicit human confirmation before running delete_instance.py against any task_id or deploy_name.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (21)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

声明描述的是一个覆盖创建、查询、删除、获取 app_id 等多项实例管理能力的技能;但当前代码片段只包含 create_instance.py,且行为仅限于创建 GPU 容器实例并处理幂等返回与错误情况。它没有展示任何查询实例状态、删除实例、清理资源或获取 app_id 的实现。因此,实际代码能力明显小于且不完全符合声明的技能用途,构成描述与行为不一致。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

声明描述的是一个“GPU云平台实例管理”技能,重点在实例的创建、查询状态、获取 app_id、删除资源等生命周期操作。但给出的代码块只包含 list_apps.py,其唯一功能是读取环境变量中的 token,并向 https://api.vylai.com/app/app 发起 GET 请求获取应用列表。返回结果也只是应用元数据(id、名称、描述、owner、gpu_num、price),没有任何实例创建、实例状态查询、实例删除或资源清理逻辑。因此,代码的实际主要用途是“列出平台应用/镜像信息”,最多只与声明中的“获取应用 app_id”有部分相关,但整体上不能代表所宣称的实例管理能力,属于明显不匹配。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

声明描述的是“GPU云平台实例管理”类技能,重点应包括创建实例、查询实例状态、删除实例/清理资源、获取 app_id 等操作。但提供的代码只实现了一个 GPU 资源列表查询脚本,访问的是 /market/allgpus 接口,返回的是平台可用 GPU 型号、算力、显存、价格和可用数量等市场资源信息,而不是用户实例信息。代码没有任何创建、删除、实例查询或 app_id 相关逻辑。因此其实际功能与声明的主要用途存在明显不一致,属于描述与行为不匹配。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

代码的实际用途较单一:只实现了实例状态查询,包括查询全部实例和按 task_id 查询指定实例,并对结果做格式化输出。它使用的外部资源(vylai.com GPU 平台 API)与声明总体领域一致,因此不是完全无关;但声明描述的是一个包含创建、查询、删除、获取 app_id 等多项操作的管理技能,而当前提供的代码片段并未体现这些关键能力。根据评估标准,这属于描述覆盖了多项核心能力,但代码实际行为仅支持其中查询部分,存在明显的能力级别不匹配。

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
91% confidence
Finding

This endpoint enables destructive deletion based on a caller-supplied path parameter, and the skill context is instance lifecycle management where an agent may act on user input. Without strong confirmation, ownership validation, and safe parameter handling in the consuming skill, an attacker or mistaken prompt could cause unintended deletion of cloud resources and associated workloads/data.

Content

Scanner excerpt · references/api-reference.md (reported line 220)May include surrounding context.

md
### 5. 删除实例(通过task_id)

**接口**: `DELETE /api/v1/app/delete/{task_id}`

通过任务ID删除实例。

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
93% confidence
Finding

This delete-by-deploy_name endpoint is dangerous because deploy_name may be copied from status output or extracted from hostnames, making it easier for an agent to act on loosely validated identifiers. In an instance-management skill, unsafe propagation of this parameter from user-controlled or environment-derived input can lead to deletion of the wrong deployment or attacker-influenced resource targeting.

Content

Scanner excerpt · references/api-reference.md (reported line 244)May include surrounding context.

md
### 6. 删除实例(通过实例ID)

**接口**: `DELETE /api/v1/deploy/delete/{deploy_name}`

通过容器部署名称删除实例。

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
82% confidence
Finding

The skill documents use of environment variables and networked API calls but does not declare any explicit tool scope or permission boundaries. This can cause an agent runtime to grant broader-than-necessary access, making credential access and outbound requests less transparent and harder to govern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The natural-language content of the skill is entirely in Chinese, including the description and operational guidance, which effectively forces a specific language for users. There is no indication that this is a region-specific skill or that users may opt into another language.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill documents deletion of cloud instances but does not prominently warn that deletion is destructive and may affect availability, running workloads, or billing. In an agent context, ambiguous destructive operations can lead to accidental resource loss or service disruption if triggered without strong confirmation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This markdown file describes two DELETE endpoints for removing instances, but it does not provide a clear user-facing warning that these operations are destructive and may terminate running workloads or affect user data/system state. Under the markdown-specific warning rule, destructive behaviors should be explicitly disclosed rather than only inferred from the endpoint names.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The script's docstrings, argument descriptions, and error/help messages are presented exclusively in Chinese, which imposes a specific language on users. Under the policy, locale or language constraints should be optional, user-selected, or explicitly justified.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
80% 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 · scripts/create_instance.py (reported line 65)May include surrounding context.

python
payload["accessory"] = accessory
    
    try:
        response = requests.post(url, headers=headers, json=payload, timeout=30)
        
        if response.status_code >= 400:
            return {

External Transmission

Medium
Category
Data Exfiltration
Confidence
80% 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 · scripts/query_instances.py (reported line 91)May include surrounding context.

python
payload["accessory"] = accessory
    
    try:
        response = requests.post(url, headers=headers, json=payload, timeout=30)
        
        if response.status_code >= 400:
            return {

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This function performs an irreversible remote deletion as soon as it is invoked, with no secondary confirmation, dry-run mode, or explicit acknowledgment requirement. In an agent skill context, where parameters may be derived from user prompts, tool chains, or mistaken automation, this increases the risk of accidental or unauthorized resource destruction.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This alternate delete path has the same unsafe behavior: it issues a destructive API call directly from a supplied deploy_name without any guardrail or confirmation. Because deploy_name may be easier for operators or agents to confuse than a task-specific workflow token, the chance of deleting the wrong instance is materially increased in automated use.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest says the skill is for creating, querying, and deleting GPU cloud instances, plus obtaining an application's app_id when needed. This file implements a broader app marketplace/listing function that enumerates platform applications, including public and private apps, which is a distinct capability not described as part of instance management.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The manifest says the skill is for creating, querying, and deleting platform instances, plus obtaining app_id and cleaning resources. This file instead queries the marketplace-wide GPU catalog via /market/allgpus, which is inventory discovery rather than instance management or instance-status lookup.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill instructs users to set an API token and notes that token-authenticated requests can access private applications, but it does not provide clear credential-handling guidance. This increases the risk of accidental exposure of secrets or unintended access to private account data in logs, prompts, or misconfigured environments.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

The entire skill file is presented only in Chinese and does not indicate any user choice of language or locale. The policy requires flagging natural-language content that forces a specific language without opt-in unless the locale restriction is clearly documented and justified.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

This Python file makes a network POST request to a remote service using an API token and sends app/container creation parameters in the request body. Although the function names and docstrings describe creation behavior, there is no explicit runtime warning, confirmation, or user-facing notice that data will be transmitted to an external API.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

This code conditionally reads an API token from environment variables and sends it in the X-Auth-Token header during an HTTP request. Although the docstring explains the functional difference between token and no-token modes, there is no user-facing prompt, warning, or log statement at execution time indicating that private-account context may be used and transmitted to the remote API.

Content

No source excerpt is available for this finding.