Back to skill

Security audit

爽文体验馆

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a game launcher skill, but it needs review because its install and save behavior is broader and less clearly bounded than users would expect.

Install only if you are comfortable with a Chinese-language game hub that may guide the agent to install additional OpenClaw skills. Before approving installation, verify the exact package names, publisher, versions, and source, and avoid the global unpinned npx commands. Treat saved game text as persistent data and prefer dedicated save files over shared agent memory.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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)

T08 · Insecure Dependencies

Error
Location
scripts/launcher.py:21
Finding

Unpinned Third-Party Skill Installation Allows Supply-Chain Code Execution

Content
View full analysis

Vulnerability Details

File Location: scripts/launcher.py:21-29; related installation instructions in references/core_logic.md:38-54 and trust assertion in references/ui_templates.md:62-69
Vulnerability Type: Unverified and unpinned third-party dependency installation
Risk Level: High

Vulnerable Code

scripts/launcher.py:21-29:

python
def install_skill(skill_name):
    """Install the specified skill."""
    print(f"Installing {skill_name}...")
    # Install through the OpenClaw CLI
    cmd = f"openclaw skills install {skill_name}"
    try:
        subprocess.run(cmd, shell=True, check=True)
        return True
    except subprocess.CalledProcessError:
        return False

references/core_logic.md:38-54:

text
Installation commands:

Cultivation game:
npx skills add cultivation-chronicle-cn -g -y

or:

openclaw skills install cultivation-chronicle-cn

Wuxia game:
npx skills add wuxia-quwen -g -y

or:

openclaw skills install wuxia-quwen

references/ui_templates.md:62-69:

text
Package information:
- Name: Cultivation Chronicle (cultivation-chronicle)
- Type: Cultivation text adventure
- Size: Approximately 2 MB
- Source: skillhub (trusted)

Install now? Reply "yes" to confirm.

Technical Analysis

The Skill installs third-party packages by mutable package name without a pinned version, integrity hash, signature check, fixed registry URL, or publisher verification. The documented npx ... -g -y alternative also performs unattended global installation and can execute package lifecycle code.

The UI labels an unspecified SkillHub source as trusted, but the implementation does not authenticate that source or establish that the resolved package is the expected reviewed artifact. The package identifiers are also inconsistent: the UI refers to cultivation-chronicle, while the implementation installs cultivation-chronicle-cn. Th ...[truncated 1830 chars]

Remediation
View remediation

Remediation Suggestions

  1. Pin every dependency to a specific reviewed version or immutable content digest.

  2. Configure an explicit approved registry and verify publisher identity, package signatures, and cryptographic hashes before installation.

  3. Remove the unattended global npx ... -g -y installation path.

  4. Present the exact package name, version, publisher, source URL, permissions, and integrity digest before requesting installation approval.

  5. Replace shell execution with an argument array:

    python
    subprocess.run(
        ["openclaw", "skills", "install", verified_skill_name],
        check=True,
        shell=False,
    )
    
  6. Enforce an internal allowlist mapping game identifiers to immutable package metadata rather than accepting arbitrary package names.

  7. Install dependencies in a restricted environment with minimum filesystem and network privileges.

  8. Reconcile the inconsistent cultivation-chronicle and cultivation-chronicle-cn identifiers.

  9. Do not describe a source as trusted unless the implementation performs corresponding authenticity and integrity checks.

T02 · Agent Memory Poisoning

Warning
Location
references/save_system.md:23
Finding

Game State Is Directed into Shared Long-Term Agent Memory

Content
View full analysis

Vulnerability Details

File Location: references/save_system.md:23-37
Vulnerability Type: Unsafe persistence of user-controlled game state in shared Agent memory
Risk Level: Medium

Vulnerable Instructions

text
Cultivation Chronicle:

workspace/
├── cultivation_save.md
├── memory/
│   └── YYYY-MM-DD.md
└── MEMORY.md

Wuxia Quwen:

workspace/
├── wuxia_save.md
├── memory/
│   └── YYYY-MM-DD.md
└── MEMORY.md

Technical Analysis

The save-system instructions place both games' records in the general memory/ directory and shared MEMORY.md long-term memory, even though the Skill repeatedly claims that game saves are completely isolated.

Narrative state can contain user-controlled names, text, choices, or instructions. Persisting that content in general Agent memory creates a cross-session influence channel: later Agent sessions may interpret saved narrative text as relevant context rather than inert game data. It also allows data from one game to influence another game or an unrelated future task.

The reviewed files do not explicitly instruct the Agent to overwrite safety rules, and no concrete malicious memory payload is embedded in the project. The risk arises from directing variable, user-influenced game content into a shared long-term memory mechanism instead of a dedicated structured save location.

Attack Path

  1. A user enters crafted text while playing, such as a character name or narrative response containing instruction-like content.
  2. The game includes that content in its current state or activity record.
  3. Following save_system.md, the Agent writes the state into memory/YYYY-MM-DD.md or shared MEMORY.md.
  4. A later session loads or consults the shared memory.
  5. The persisted game text is exposed outside its original game context and may be treated as Agent-relevant context.
  6. This can contaminate unrelated responses or violate the promised separati ...[truncated 537 chars]
Remediation
View remediation

Remediation Suggestions

  1. Never store game state in general-purpose MEMORY.md or the Agent's shared daily-memory directory.

  2. Create a dedicated save namespace for each game, for example:

    text
    workspace/game-saves/cultivation/state.json
    workspace/game-saves/wuxia/state.json
    workspace/game-saves/minguo/state.json
    workspace/game-saves/campus/state.json
    
  3. Store state as schema-validated structured data rather than executable or instruction-like Markdown.

  4. Treat all loaded save fields as untrusted data and clearly delimit them from Agent instructions.

  5. Apply length limits and reject control fields or unexpected properties.

  6. Ensure one game's loader cannot read another game's save directory.

  7. Obtain explicit consent before retaining user-provided narrative data across sessions and provide a deletion mechanism.

  8. Update the documentation so its isolation claims accurately match the implemented storage model.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (26)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description says this skill is a four-world entry hub covering 修仙、武侠、民国、校园, with independent saves for each. The supplied code only supports two options: 凡人踏天行 and 武林趣闻. It has no menu entries, details, or installation checks for the 民国 or 校园 worlds. It also checks the filesystem under ~/.openclaw/workspace/skills/ to determine whether two subordinate skills are installed, which is an actual behavior not reflected in the description. Finally, the claim that the four worlds have separate save data is not substantiated by this code; there is no save management logic at all. Therefore the description materially overstates and misrepresents the implemented behavior.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The description promises a four-world hub with entries for 修仙、武侠、民国、校园 and states that all four have separate saves. The supplied code only presents two options: 凡人踏天行 and 武林趣闻. It does not include the 民国 or 校园 worlds at all. Additionally, the code accesses local filesystem paths under ~/.openclaw/workspace/skills to detect whether the two worlds are installed, which is an undeclared behavior/resource access. There is also no implementation here for independent save handling. The primary purpose is related—a welcome/selection screen—but the delivered behavior materially underimplements and partially differs from the declared functionality.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

声明的核心功能是“四世界统一入口”,包括修仙、武侠、民国、校园四个独立世界及独立存档。实际代码中只定义了两个技能名 cultivation-chronicle-cn 和 wuxia-quwen,欢迎文本也仅列出 1 和 2 两个选项,没有民国或校园相关内容。代码还包含 check_skill 和 install_skill,用于检查本地技能目录并通过 openclaw CLI 安装技能,这属于未在描述中说明的能力。与此同时,main() 只是打印欢迎文本,并没有真正处理用户输入、进入世界或管理存档。因此,实际行为与声明用途存在明显且实质性的偏差。

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
97% confidence
Finding

This is a tool-parameter abuse pattern: the subprocess invocation delegates command construction to the shell and includes a variable parameter. In a skill context, that is especially risky because tool execution may be triggered by higher-level orchestration or later code changes, allowing attacker-controlled values to escape into shell syntax and execute unintended commands.

Content

Scanner excerpt · scripts/launcher.py (reported line 26)May include surrounding context.

python
# 这里使用 openclaw CLI 安装
    cmd = f"openclaw skills install {skill_name}"
    try:
        subprocess.run(cmd, shell=True, check=True)
        return True
    except subprocess.CalledProcessError:
        return False

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.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill description is entirely in Chinese and presents the experience as a fixed Chinese-language interface, with no indication that users may choose another language. Under the policy, forcing a specific language without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The activation phrases and in-game commands are specified only in Chinese, which constrains use to a single language without stating that this is optional or region-specific. This can violate language/locale policy when no user choice or justification is provided.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The manifest describes a unified entry for four independent worlds: 修仙、武侠、民国风云、青春校园. In code, the launcher only checks installation status for two skills and only presents options 1 and 2, so the implemented behavior does not match the advertised scope of the skill.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The manifest describes a unified selection interface for four worlds: 修仙、武侠、民国风云、青春校园, each with independent saves. However, this file's selection, installation, and launch logic only covers two skills: cultivation-chronicle-cn and wuxia-quwen. That is a direct mismatch between the stated skill scope and the documented behavior.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The markdown normalizes global, non-interactive installation commands and only asks a minimal 'install now?' confirmation, without clearly warning that these commands modify the user's environment. In this skill context, users may treat the instructions as safe gameplay setup, making them more likely to approve system-changing operations without understanding the risk.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The skill instructs use of npx skills add ... -g -y without pinning a version, which can fetch and execute whatever package version is current at install time. In a skill launcher context, this creates a supply-chain risk where a compromised or typosquatted package, or a later malicious release, could execute code on the user's system during installation.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

This second npx skills add ... -g -y command has the same issue: it relies on an unpinned package and performs unattended global installation. That combination increases the blast radius of any supply-chain compromise because arbitrary code may run with the privileges of the installing user and persist system-wide.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

清单描述称该技能提供修仙、武侠、民国、校园四个世界入口,且“四个世界独立存档,互不影响”。但本文件仅为“凡人踏天行”和“武林趣闻”定义了独立原则、文件位置和切换流程,没有覆盖“民国风云”和“青春校园”的存档设计,实际行为范围明显少于描述承诺。

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The file's user-facing comments and printed interface are entirely in Chinese, with no indication that the user can choose another language or that the skill is intentionally limited to a Chinese-only audience. This is a natural-language policy concern because it imposes a locale/language constraint implicitly rather than offering opt-in or justification.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Manifest 描述该技能应提供修仙、武侠、民国、校园四个世界的统一角色选择入口,且四个世界独立存档。此文件实际只展示并允许选择 1 和 2 两个世界,没有民国风云或青春校园入口,行为范围明显小于描述承诺。

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The top-level documentation states '自动检测、安装并启动修仙或武侠游戏', implying the script will detect installation state and perform install/start actions. In practice, main() only prints welcome text, while install_skill(), start_xiuxian(), and start_wuxia() are never invoked, so the documentation overstates what the script actually does.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The manifest describes a unified role-selection interface for four worlds: 修仙、武侠、民国、校园, with independent saves for all four. In this file, only two skills are configured and presented to the user, and there is no code for 民国 or 校园 world launch paths or any save-isolation handling, so the implemented behavior falls materially short of the stated skill description.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The code contains an installation routine that can fetch and install external skills without any confirmation, trust prompt, or warning. In a skill ecosystem, installing additional packages or content expands the trust boundary and can expose users to unreviewed code or malicious skill content if this function is later wired into the UI.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
95% confidence
Finding

The launcher builds a shell command with string interpolation and executes it with shell=True. Although the current callers pass hardcoded skill names, this pattern is dangerous because any future untrusted input reaching skill_name could trigger command injection and arbitrary command execution in the user's environment.

Content

Scanner excerpt · scripts/launcher.py (reported line 26)May include surrounding context.

python
# 这里使用 openclaw CLI 安装
    cmd = f"openclaw skills install {skill_name}"
    try:
        subprocess.run(cmd, shell=True, check=True)
        return True
    except subprocess.CalledProcessError:
        return False

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

All user-facing natural-language strings in the file, including the title, banner, menu, and detailed descriptions, are Chinese-only. This can violate language/locale policy when a skill forces a specific language without user opt-in and no region-specific justification is provided.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

The inline user-facing documentation in the welcome banner states that the player can choose '两种截然不同的人生', which conflicts with the skill's stated purpose of providing access to four distinct worlds. This is an intent/documentation contradiction rather than a minor omission because it changes the advertised scope of the launcher.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The welcome text explicitly states the player can choose '两种截然不同的人生', and the menu below presents only options 1 and 2. This conflicts with the skill's declared intent of providing four world entrances, making the embedded documentation inconsistent with the stated purpose.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

This is a markdown file, so SQP-2 applies to omitted warnings about behaviors affecting user data. The text states that progress is '自动存档' and identifies save-file locations, but it does not explicitly warn users that switching games will cause automatic writes to workspace save files.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown file's user-facing templates are entirely in Chinese, and there is no indication that users can opt into another language or that the skill is intentionally limited to a Chinese-speaking audience. Under the policy, forcing a specific language without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

L25 的用户提示文案明确写出“你可以选择两种截然不同的人生”,这与技能说明中“四世界入口”的定位直接冲突。这里不是单纯信息缺失,而是界面文案主动将技能能力表述为两种选择,和对外声明形成明显矛盾。

Content

No source excerpt is available for this finding.