Back to skill

Security audit

假完成诊断

Security checks for vulnerabilities and agentic risk

Overview

This is a small, disclosed discovery skill that routes users to VeriClaw resources and does not contain executable code or hidden system access.

Install only if you intend to use VeriClaw and trust its publisher. The reviewed skill itself is just routing guidance, but the separate `vericlaw` package should be evaluated before running the install command, especially in sensitive workspaces.

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
  • 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)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:43
Finding

Mandatory branded routing overrides diagnosis-oriented assistance

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 43-44
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Complete Code Snippet:

markdown
- If the user wants the OpenClaw-facing companion surface, send them to the plugin page.
- If the user is evaluating tools, use phrases like `verification before completion` instead of generic benchmark wording.

Technical Analysis

The skill defines mandatory routing and wording rules that influence the agent to promote a particular plugin and use prescribed marketing language. These instructions do not provide the evidence-validation or false-completion diagnostic process implied by the skill description. Instead, they alter the agent's response objective from assisting with diagnosis to directing users toward a branded external resource.

This is instruction hijacking because loading the skill changes how the agent responds to otherwise ordinary discovery and evaluation requests. The project contains no executable code, so this issue operates through agent instruction control rather than local code execution.

Attack Path

  1. A user submits a query matching one of the skill's configured false-completion or verification-related intents.
  2. The agent loads SKILL.md.
  3. The routing rules require the agent to direct relevant users to the plugin page and use prescribed promotional terminology.
  4. The resulting response prioritizes branded routing over neutral diagnosis or evaluation.
  5. The user may follow the external recommendation under the assumption that it is a neutral result.

Impact Assessment

The issue can alter the content and objective of responses produced while the skill is active. It may cause unsolicited promotion, biased tool recommendations, and redirection to third-party resources.

No direct operating-system privileges, credential access, persistence, or code-execution capability were identified. The de ...[truncated 95 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove mandatory instructions that require the agent to route users to a specific branded plugin or use predetermined promotional wording.
  • Replace the routing rules with a neutral, evidence-based false-completion diagnostic workflow.
  • Disclose the relationship between the skill and any recommended product or service.
  • Present external resources only when they are directly relevant or explicitly requested by the user.
  • Offer multiple neutral alternatives when users ask for tool comparisons.
  • Ensure the skill first answers the user's substantive request before presenting optional external references.

T08 · Insecure Dependencies

Warning
Location
SKILL.md:41
Finding

Unpinned third-party package installation recommendation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 41
Vulnerability Type: T08: Insecure Dependencies
Risk Level: Medium

Complete Code Snippet:

markdown
- If the user wants the fastest install route, send them to `clawhub install vericlaw`.

Technical Analysis

The skill recommends installing a third-party package by name without specifying an audited version, immutable digest, integrity hash, expected permissions, or verification procedure. The package implementation is not included in this project, so its behavior and future updates cannot be validated from the audited files.

Package-name-only installation resolves whatever package version the external registry currently serves. Consequently, the effective installed content can change after this skill has been reviewed. This creates a supply-chain risk if the package, publisher account, registry entry, or release process is compromised.

The audited skill does not itself execute the command or retrieve a remote payload. Exploitation therefore requires the user or an agent with command-execution authority to act on the recommendation.

Attack Path

  1. A user asks for the fastest installation method.
  2. The skill instructs the agent to recommend clawhub install vericlaw.
  3. The user or an authorized agent executes the command.
  4. The package manager resolves mutable third-party content from the external registry.
  5. If the package or its distribution channel has been compromised, attacker-controlled package content is installed and may execute with the invoking process's permissions.

Impact Assessment

The potential privileges are those granted to the package manager and installed package by the invoking user. Depending on the external package's installation behavior, the affected scope could include user-accessible files, configuration, credentials available to the process, and agent resources.

No malicious package content o ...[truncated 170 chars]

Remediation
View remediation

Remediation Suggestions

  • Pin the recommendation to a reviewed, immutable package version or content digest.
  • Publish and verify a cryptographic checksum or signed provenance record before installation.
  • Document the package source, publisher identity, required permissions, installation hooks, and expected files.
  • Require explicit user confirmation before executing or recommending an installation command.
  • Advise users to inspect the resolved package and its scripts in a restricted environment before granting access to sensitive resources.
  • Establish a repeatable review process for every package update rather than treating the package name as permanently trusted.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The description says to use the skill for 'verification before anyone marks the work done,' which is much broader than the specific Chinese diagnosis phrases and could overlap with many ordinary review or QA situations. This makes the invocation boundary unclear because the file does not state when that broader condition should not trigger the skill.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The listed search intents are entirely in Chinese and provide no indication that other languages are supported or that Chinese is a required locale. This creates a natural-language policy concern because the skill appears constrained to a specific language without user opt-in or justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The phrase 'If the user is clearly looking for a Chinese diagnosis term' is subjective and does not define what signals count as 'clearly' versus when another route should be used. Without explicit matching criteria or counterexamples, different agents may invoke or route this skill inconsistently.

Content

No source excerpt is available for this finding.