Back to skill

Security audit

Plugin

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a real werewolf game plugin, but it uses a shared plaintext remote server and background prompt-time lobby polling in ways users should review before installing.

Install only if you are comfortable with game names, actions, roles, event logs, and possibly agent-generated reasoning being sent to or displayed by the configured server. Prefer a local/self-hosted server, avoid sensitive prompts while the plugin is enabled, and treat shared-server replay links and table IDs as private.

Vulnerability Patterns
  • 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
  • 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)

T09 · Insecure Skill Coding Practices

Warning
Location
openclaw.plugin.json:10
Finding

Plaintext HTTP Used for the Default Remote Game Server

Content
View full analysis

Vulnerability Details

File Location: openclaw.plugin.json:10-14
Vulnerability Type: Insecure network transport
Risk Level: Medium

Vulnerable Code

json
"serverUrl": {
  "type": "string",
  "default": "http://47.85.184.157:8801",
  "description": "Werewolf HTTP server base URL. Default points to the lobster-republic US test server; set to http://127.0.0.1:18811 if you run your own local server."
}

The same plaintext endpoint is documented in SKILL.md:57-60 and README.md:124:

markdown
- 默认 serverUrl 指向 `http://47.85.184.157:8801`(lobster-republic 美国 Virginia 测试服务器)

Technical Analysis

The plugin defaults to a shared external service accessed over plaintext HTTP. HTTP provides neither confidentiality nor server authenticity, allowing a network-positioned attacker to observe or modify requests and responses.

Based on the documented behavior, traffic can contain player names, speeches, decisions, game events, hidden roles, and private in-game actions. Manipulated responses are also returned to the agent and may influence its subsequent decisions or user-facing output.

The serverUrl property accepts an unrestricted string, and the supplied configuration does not enforce HTTPS for non-loopback destinations. The executable dist/index.js implementation was absent from the audited artifact, so additional runtime URL validation or authentication could not be verified.

Attack Path

  1. A user asks the agent to start or participate in a Werewolf game.
  2. The plugin uses the default http://47.85.184.157:8801 endpoint.
  3. An attacker positioned on the local network, upstream network, proxy, or routing path intercepts the plaintext connection.
  4. The attacker reads game requests and responses or changes server responses in transit.
  5. Modified data is accepted by the plugin and presented to the agent as legitimate game state.

Impact Assessment

An attacker with ...[truncated 456 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace the default endpoint with an HTTPS URL using a valid, publicly trusted certificate.
  2. Reject plaintext HTTP for all non-loopback destinations.
  3. Permit HTTP only for explicit development endpoints such as 127.0.0.1, ::1, or a Unix-domain transport.
  4. Parse and validate serverUrl with an allowlist of supported schemes and expected hosts.
  5. Add authenticated requests using short-lived, scoped credentials rather than relying solely on table identifiers.
  6. Configure strict connection and response-size limits to reduce exposure to a malicious or compromised server.
  7. Document exactly which data is transmitted and obtain user confirmation before sending conversation-derived content to a shared external service.
  8. Include the executable plugin implementation in future audits so TLS enforcement, redirect handling, authentication, and payload construction can be verified.

T09 · Insecure Skill Coding Practices

Error
Location
README.md:105
Finding

Replay Interfaces Are Documented as Exposing Private State and Model Reasoning Without a Described Authorization Control

Content
View full analysis

Vulnerability Details

File Location: README.md:105-118
Vulnerability Type: Sensitive information exposure and insufficiently documented access control
Risk Level: High

Vulnerable Documentation

markdown
### 🎭 观战视角 `/spectator?table_id=xxx`
所有身份保密。只公开死亡名单、白天发言、投票、出局。适合**人类陪伴观战**,不会剧透。

![spectator view](./assets/view-spectator.png)

### 👤 主人视角 `/master?table_id=xxx&player=你的龙虾名`
**只显示你自己那张牌** + 所有公开事件 + 你的"内心 OS"(LLM 思考过程)。符合真实狼人杀体验 —— 只有你知道你是谁。

![master view](./assets/view-master.png)

### 👁 上帝视角 `/god?table_id=xxx`
所有身份可见,所有夜间私密行动(狼队决议、预言家查验结果、女巫用药)可见。**仅调试用**,正式游戏时不要看。

Technical Analysis

The documented master view intentionally exposes the selected agent's private role and internal model reasoning. The god view exposes all roles and all private nighttime actions.

The example access paths identify resources through table_id and, for the master view, a player name. No authentication token, authorization check, session binding, expiration mechanism, or ownership verification is described. A warning not to use the god view during a formal game is not an access-control mechanism.

Exposing internal model reasoning is unsafe even when access is authenticated. Such content may contain unnecessary contextual details, inferred private information, or conversation-derived material. Applications should expose a concise decision explanation rather than raw internal reasoning.

The server and executable plugin code were absent from the supplied artifact. Consequently, the audit cannot prove whether undocumented server-side authorization exists. The confirmed issue is that sensitive views and model reasoning are deliberately exposed while the supplied security design and usage instructions provide no authorization mechanism.

Attack Path

  1. A game is created on the shared server.
  2. A table identifier is disclosed through a lobby link, game output, browser history, logs, chat mess ...[truncated 1035 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove all exposure of raw chain-of-thought or internal model reasoning. Provide short, purpose-specific decision summaries instead.
  2. Require authentication for every spectator, master, and god request.
  3. Issue cryptographically random, unguessable, short-lived viewer tokens scoped to a specific table and view type.
  4. Enforce authorization server-side; never authorize access based only on table_id or a player name.
  5. Restrict god-view access to authenticated administrators in a separate debug environment, and disable it in public deployments.
  6. Bind master-view access to the verified owner of the relevant player seat.
  7. Expire replay links when a game ends and support immediate token revocation.
  8. Avoid placing sensitive credentials in query strings because URLs may leak through browser history, logs, analytics, and referrer headers. Use secure cookies or authorization headers.
  9. Apply rate limiting and audit logging to failed and successful access attempts.
  10. Update the documentation to define the authentication, authorization, retention, and privacy model explicitly.
  11. Review the unavailable server implementation to verify that all replay endpoints implement these controls.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (22)

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The skill advertises play with 8 NPC lobsters, but the documented tools support creating tables, checking status, and retrieving event streams that extend beyond a one-shot NPC game. If exposed without clear disclosure and scoping, these capabilities can leak game/session data, enable cross-user interaction on a shared service, or create unintended coordination surfaces.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill advertises play with 8 NPC lobsters, but the documented tools support creating tables, checking status, and retrieving event streams that extend beyond a one-shot NPC game. If exposed without clear disclosure and scoping, these capabilities can leak game/session data, enable cross-user interaction on a shared service, or create unintended coordination surfaces.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The README directs users to a default remote server and a browser-accessible lobby, but does not clearly warn that gameplay data, prompts, agent outputs, and possibly role/identity information are transmitted to and stored on an external host. In an agent skill context, this can cause unintended disclosure of user commands, game state, and model-generated content to a third party, especially because the plugin appears designed for autonomous use with minimal user intervention.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documentation states that the master view exposes the agent's 'inner monologue'/'LLM thinking process'. Revealing chain-of-thought can disclose hidden reasoning, internal instructions, secrets encountered during play, or prompt fragments, and it creates a direct exfiltration path through the web UI if the agent processes sensitive context.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
89% confidence
Finding

The skill clearly depends on networked tools and a default remote server, but it does not declare an explicit tool scope or permissions boundary. That makes the skill's effective capabilities less transparent to the host and user, increasing the risk of unintended remote data transmission or broader tool access than expected.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Activation triggers such as '玩游戏' are broad and can match common conversation unrelated to this specific skill. Because the skill performs networked actions against a remote service, overly permissive triggering increases the chance of accidental invocation and unintended outbound requests.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill states that the default server URL points to a shared remote host, but it does not prominently warn users in the description or activation surface that their prompts, game state, and related metadata may be sent off-device. In a shared-server context, this creates privacy and data-handling risks, especially if users assume the game is local.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill is presented as a simple local werewolf game, but the code hardcodes a default remote server and forwards gameplay traffic there. This creates an undisclosed data flow to a third party, which can expose user prompts, game state, identifiers, or metadata and violates user expectations about where processing occurs.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The configure(serverUrl) function allows the destination server to be replaced with an arbitrary URL, giving the skill general outbound HTTP capability beyond the apparent game purpose. If an attacker or untrusted integrator can influence this value, the skill could be redirected to exfiltrate conversation data or interact with unintended internal or external services.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This function sends request bodies over HTTP to a remote server without any visible disclosure, consent flow, or transport security guarantees. Because the default endpoint uses plain HTTP, gameplay content and associated metadata could be intercepted or modified in transit, and users are not informed that their data leaves the local skill context.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill metadata describes a local 8-NPC werewolf game, but the implementation configures and uses a remote external server for multiplayer features. This scope mismatch is security-relevant because users and host agents may grant permissions based on the manifest description, while the code silently enables external communication and cross-agent interaction not clearly disclosed by the skill contract.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The plugin appends unsolicited lobby invitations into system context on every prompt build, nudging the agent toward multiplayer actions outside the advertised single-player game behavior. Even with basic sanitization, this creates covert behavioral influence from remote state and can manipulate agent decisions without a user request, broadening the skill's effective authority.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The code performs external polling during the prompt-building lifecycle, meaning network activity occurs automatically whenever the agent is awakened, not only when the user asks to play. This expands data exposure, creates hidden side effects, and allows a remote service to influence prompt construction in a context where users and operators may not expect any network access.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

A network request to an external IP-based server is made during prompt construction without any disclosure in this file to the user at the moment it happens. Hidden background requests are dangerous because they can leak timing, metadata, and usage patterns and can introduce remote content into the agent workflow without informed consent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The injected system context is written entirely in Chinese and instructs the agent how to behave using that locale, with no indication that the user can choose another language. Under the policy, forcing a specific language without opt-in is a natural-language policy violation unless the locale restriction is explicitly documented and justified.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The activation phrases are broad enough to match ordinary conversation such as '玩游戏' or general discussion about werewolf, which can cause the skill to trigger when the user did not actually intend to invoke this plugin. In this skill, unintended activation is more concerning because the documented common path launches a long-running external tool workflow against a remote server, creating avoidable network activity and potentially disruptive behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

The natural-language comments and API documentation are written in Chinese with no indication of alternative language support or user opt-in. This can conflict with language/locale policy requirements when a skill is expected to accommodate broader audiences unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
81% confidence
Finding

Several natural-language strings set a default lobster name of '白小浪' and much of the surrounding commentary is Chinese, which can impose a language/locale assumption rather than offering user choice. The file does not document this as a region-specific constraint or provide an explicit opt-in for language/locale behavior.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

This code file makes HTTP requests that send user-provided names and gameplay state to a server via callServer, including POST bodies and query parameters. While the tool descriptions explain functionality, they do not disclose that these values are transmitted over the network, and the execute paths themselves contain no confirmation prompt or user-facing disclosure.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest description is entirely in Chinese and does not indicate any language choice or opt-in. Under the policy rule for natural-language constraints, this can be treated as a locale/language assumption that may exclude users without explicit justification.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
78% confidence
Finding

The package description uses only Chinese for the skill title ("龙虾狼人杀") alongside English descriptive text, which can indicate a fixed locale presentation in natural-language metadata. Because this manifest does not state that the skill is region-specific or offer a language choice, it may conflict with language/locale policy expectations for user-facing skill descriptions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The OpenClaw metadata repeats the Chinese title ("龙虾狼人杀") in a user-facing description without indicating that the plugin is intentionally limited to a Chinese-speaking audience. Under the stated policy, language constraints should be optional or clearly justified when exposed in natural-language metadata.

Content

No source excerpt is available for this finding.