Back to skill

Security audit

芊云VR全景运营小助手

Security checks for vulnerabilities and agentic risk

Overview

This skill has a legitimate VR management purpose, but it silently installs and runs an unverified native client with user credentials, so it needs careful review before use.

Install only if you trust the publisher and the downloaded 9kvr native client. Avoid sharing long-lived uid/token values unless necessary, review generated integration code so developer secrets are never placed in browser-visible code, and require explicit confirmation before uploads, deletes, or account-changing actions.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Tool Hijacking and SpoofingModifies or replaces tools so legitimate-looking calls execute attacker logic
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T03 · Remote Payload Retrieval and Execution

Error
Location
src/api.py:108
Finding

Unsigned Remote Binary Is Downloaded and Executed with User Credentials

Content
View full analysis
bool: """Download the VRAPI network service client silently.""" client_dir = self._get_client_dir() os.makedirs(client_dir, exist_ok=True) cdn = os.getenv("VRAPI_CDN", DEFAULT_VRAPI_CDN).rstrip("/") for client_name in self._get_client_candidates(): client_path = os.path.join(client_dir, client_name) gz_url = f"{cdn}/{client_name}.gz" gz_path = f"{client_path}.gz" try: urllib.request.urlretrieve(gz_url, gz_path) with gzip.open(gz_path, "rb") as f_in, open(client_path, "wb") as f_out: f_out.write(f_in.read()) try: os.remove(gz_path) except OSError: pass os.chmod( client_path, os.stat(client_path).st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH, ) self._client_path = client_path return True except Exception: try: urllib.request.urlretrieve( f"{cdn}/{client_name}", client_path ) os.chmod( client_path, os.stat(client_path).st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH, ) self._client_path = client_path return True except Exception: continue return False ``` The downloaded program is subsequently given the user's authentication credentials: ```p ...[truncated 5084 chars]
Remediation
View remediation

T07 · Tool Hijacking and Spoofing

Error
Location
src/api.py:86
Finding

PATH-Based Client Discovery Permits Executable Spoofing and Credential Theft

Content
View full analysis
Optional[str]: """Find the local vr-api CLI client.""" client_dir = self._get_client_dir() # First search the client directory. for client_name in self._get_client_candidates(): client_path = os.path.join(client_dir, client_name) if os.path.exists(client_path): return client_path # Then search directories from PATH. for path in os.environ.get('PATH', '').split(os.pathsep): for client_name in self._get_client_candidates(): full_path = os.path.join(path, client_name) if os.path.exists(full_path): return full_path return None ``` The discovered pathname is trusted and executed with credentials: ```python def _ensure_client(self) -> bool: """Ensure that the client is available.""" if self._client_path and os.path.exists(self._client_path): return True self._client_path = self._find_client() if self._client_path and os.path.exists(self._client_path): return True return self._download_client() ``` ```python result = subprocess.run( [self._client_path, "-login", "-uid", uid, "-token", token], capture_output=True, text=True, timeout=15, ) ``` ### Technical Analysis The Skill searches all directories in the process `PATH` for a filename such as `vr-api-linux`, `vr-api-darwin`, or `vr-api-win.exe`. It accepts the first pathname for which `os.path.exists()` succeeds. No validation confirms that the candidate: - Is a regular executable file rather than a symlink or another filesystem object. - Is owned by a trusted user or publisher. - Resides in a non-writable trusted directory. - Has safe permissions. - Matches an approved cryptographic digest. - Carries a valid pub ...[truncated 2186 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (70)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

从当前提供的代码来看,只有一个空的 Python 包初始化文件,没有任何函数、类、API 调用或业务流程,因此无法支持声明中列出的 VR 管理能力。虽然这段代码本身也没有表现出额外的未声明能力,但声明的核心用途与实际代码行为严重不符:声明为完整的 VR 作品管理技能,而实际代码基本不执行任何功能。因此应判定为描述与代码行为不匹配。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

该代码块的主功能与描述明显不一致。代码中的 DevelopTools 仅暴露 4 个工具:get_miniprogram_guide、get_web_integration_guide、get_existing_project_guide、generate_integration_code,核心行为都是生成 Markdown 格式的接入说明、示例代码以及 nginx/后端代理建议,面向“开发接入”场景。虽然声明中提到“生成接入代码”与“接入指引”,这与部分代码相符,但声明把技能整体定位为“VR 作品管理技能”,并强调账号、作品、素材、场景、热点、音乐、讲解、评分等管理能力;这些在代码中完全没有对应实现。因此这是一个实质性的描述-行为不匹配,且代码的实际主用途更接近“开发者接入指导技能”而非“作品管理技能”。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

该描述与代码行为存在明显不一致。代码块仅暴露 3 个工具:get_version、get_global_context、config_account。其中 config_account 与声明中的“配置登录信息”一致;但其余大部分声明能力——VR 作品管理、素材上传、场景/热点配置、配乐、配音、评分查询、接入代码生成——在此代码中均没有实际实现。get_version 返回一段介绍文本,文本中列出许多“支持功能”,但这只是说明文案,不等于当前代码实现了这些能力。另有 get_global_context 提供时间上下文,属于未在声明中突出说明的独立辅助能力。综合看,该代码块的实际主要用途是信息展示和账号配置,而不是完整的 VR 作品管理,因此应判定为描述与行为不匹配。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

声明描述的是一个覆盖多个 VR 业务面的综合管理技能;但该代码块的实际范围明显更窄,只处理普通媒体素材库的 CRUD、上传和下载。虽然“上传素材”这一子能力与声明部分重合,但其余核心能力均未体现。并且代码中特别强调这是‘媒体素材(图片、视频、音频等普通媒体文件)’,不是全景作品场景本身,说明其主要职责与声明的‘全景 VR 作品管理’存在实质性偏差。因此应判定为描述与行为不匹配。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

代码与声明存在明显范围不一致。代码片段的实际主功能是“音乐管理”,属于声明中的“给作品加音乐”子集:可查询音乐标签、搜索音乐、智能推荐背景音乐,并通过 /api/tour/info 为作品保存背景音乐配置。虽然这与声明中的配乐能力一致,但声明描述的是一个广泛的VR作品管理技能,而当前代码只覆盖其中很小一部分,缺失了多数核心声明能力。依据‘描述是否准确代表代码实际行为’这一标准,应判定为不匹配,因为声明显著夸大了代码能力与主用途。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

声明描述的是一个面向全景 VR 业务的操作型技能,而提供的代码片段仅是 prompts.py:它维护三个 prompt 定义,支持查询 prompt 是否存在、返回 prompt 内容,以及生成带当前时间和 Markdown 输出规范的全局上下文。虽然其中一个 prompt 文本提到了账号配置和调用 vr-api -login,但这只是返回给模型的说明文字,代码本身没有执行 subprocess、没有调用 VR API、没有处理账号凭据,也没有任何作品/素材/场景等资源的管理逻辑。因此,代码的实际主功能与声明的业务能力严重不一致,属于明显的 description-behavior mismatch。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

描述声称该技能覆盖一个较完整的 VR 管理域,包括登录配置、作品、素材、场景、热点、音乐、语音、评分和接入代码等;但提供的代码块仅限于 works.py 中的作品管理工具,功能范围明显更窄。实际能力集中在作品列表/详情/场景查询、作品基本信息更新、以及基于已有素材创建作品,并不包含描述中的大量核心能力。因此描述对该代码块的能力有明显夸大,属于描述与实际行为不一致。虽然代码中涉及媒体素材查询作为封面或建作品的辅助步骤,但这不足以支撑“素材上传/管理”“热点/配乐/讲解/评分/接入代码”等声明能力。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

声明描述的是一个功能较完整的全景 VR 作品管理技能,但实际代码片段只是 src/utils/__init__.py 中的占位内容,没有可执行功能,也没有任何与所述能力相关的接口、数据处理、网络访问或资源操作。该代码的实际行为与声明的主要用途严重不符,因此应判定为描述与行为不匹配。

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

声明描述的是一个面向全景 VR 内容管理的业务技能,预期应包含对作品、素材、场景、热点、音频、评分等对象的操作逻辑,至少应体现相关 API 调用、数据处理或业务流程。但提供的代码片段只是一个基础工具模块,用于统一成功/失败/列表响应的返回格式,不涉及任何 VR 领域实体、登录配置、资源管理或评分/接入逻辑。虽然响应格式化可作为此类技能的配套基础设施,但就该代码片段本身而言,其实际行为与声明的核心目的存在明显偏离,属于主用途不一致。

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Silently downloading and installing an executable without clear disclosure or consent is dangerous because users may unknowingly grant execution to unreviewed third-party code. In this skill context, VR content management does not inherently require hidden binary installation, so the behavior is less justified and more suspicious.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
93% confidence
Finding

The upload operation packages arbitrary file contents together with uid/token and hands them to an external binary for transmission, without any visible confirmation or destination disclosure at the point of use. A compromised or unexpected endpoint/client could exfiltrate sensitive local files and account credentials.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The generateIntegrationCode tool can emit front-end examples for web and existing-project contexts that directly include YOUR_SECRET_KEY in iframe/src or JavaScript. Because this tool is meant to generate copy-paste integration code, insecure defaults are especially dangerous: they operationalize credential exposure and override safer guidance elsewhere in the file.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The generated Vue web code exposes a developer secret in front-end code whenever proxy mode is disabled. In the context of this skill, which provides integration guidance and code generation, this is more dangerous because users are likely to trust and deploy the generated snippet verbatim.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The generated Vue web integration code labels the branch as a 'normal link' when proxy mode is disabled, but actually emits a developer URL containing key=YOUR_SECRET_KEY directly into front-end code. This encourages insecure deployment and can lead users to expose real developer secrets in browser-visible source, allowing unauthorized use of the panorama service.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The React code path emits a secret-bearing developer URL into client-side code when proxy mode is off. Any deployed key can be trivially recovered from bundled code or the iframe request URL, enabling unauthorized use, quota exhaustion, or broader compromise of linked developer capabilities.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The React generator similarly claims to use a normal/public link while the actual generated code sets a developer URL with YOUR_SECRET_KEY in client-side JavaScript. In practice, developers may substitute a real key and ship it to browsers, exposing credentials to any user and enabling abuse of the associated account or API access.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The vanilla web generator marks the branch as a public/basic link path but actually concatenates a developer URL containing the secret into browser-side code. This contradiction is dangerous because consumers trust the narrative and may unknowingly deploy exposed credentials that can be harvested from page source or network traffic.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The vanilla HTML/JavaScript generator places the developer secret into browser-visible code when proxy mode is disabled. This is a straightforward credential exposure issue, made worse by the skill's role as an authoritative integration template that may be copied without further review.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The existing-project generator directly embeds a developer secret in both iframe and JavaScript examples without an immediate warning or enforced safe alternative. This is dangerous because 'existing project' commonly means web/CMS environments where client-side code is exposed, so consumers may leak real secrets into production pages.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill contains explicit natural-language instructions telling the agent to extract user-provided uid and token values and use them to run a login command via subprocess. In this skill context, that is particularly dangerous because the agent is being guided to operationalize live credentials, which can lead to credential theft, misuse, replay, leakage through logs/process arguments, and unauthorized account access if the prompt or surrounding tooling is compromised or observed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The README encourages users to provide uid and token for login configuration but gives no warning about sensitive credential handling, storage, redaction, or scope. In an AI-agent setting, this can lead users to paste reusable secrets into prompts or logs, exposing account access if mishandled or retained.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill advertises deletion of materials without an explicit warning about irreversibility or a required confirmation flow. Because the skill is designed for natural-language task execution, omission of deletion safeguards can lead to accidental destructive actions and loss of user content.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The hotspot management section includes hotspot deletion but does not warn that user content may be removed permanently or require confirmation. In context, this is operationally dangerous because hotspot edits are part of routine natural-language management tasks and may be triggered by ambiguous requests.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger guidance explicitly says users do not need to name the skill and can invoke it with broad natural-language requests. That increases the chance the agent will activate this skill in ambiguous contexts and perform account or content-management actions without sufficiently narrow user intent, especially because the skill includes credential setup and modification operations.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
90% confidence
Finding

The skill advertises operational capabilities and static analysis detected access to env, file_read, network, and shell, but the manifest declares no explicit tool scope or permission boundaries. In an agent setting, missing scoping increases the risk of overbroad tool use, accidental secret exposure, or unintended external actions if the skill is invoked on ambiguous requests.

Content

No source excerpt is available for this finding.