ClawHub
Back to skill

Security audit

Corporate Banker Digital Employee

Security checks for vulnerabilities and agentic risk

Overview

This banking workflow skill is mostly framed as human-reviewed guidance, but it contains conflicting instructions that could be read as creating or submitting regulated credit cases.

Review this skill before installation if it will be used near real credit systems. Treat all API labels, saving, case creation, submission, and case-number language as human-performed institutional workflow only; do not grant an agent write or submit permissions unless Module 8 is corrected to remove the ambiguity. Use only authorized accounts and desensitized customer data.

A.I.G

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

SkillSpector

Vulnerability Patterns
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • 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 (4)

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Module 8 repeatedly states the skill only prepares a draft and that submission must be performed by authorized personnel, but later text says the skill initiates case creation and submission. In a banking workflow, this contradiction can cause an agent or integrator to treat the skill as authorized to perform a high-impact state-changing action, risking unauthorized credit-case creation, duplicate submissions, or bypass of human control points.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Workflow step 5 explicitly instructs the skill to '发起案件创建' after confirmation, which directly conflicts with the declared execution boundary that the skill does not create or submit cases. Because this step is procedural and imperative, downstream agents may execute it as operational guidance, enabling unauthorized transaction-like actions in a regulated credit system.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The example says the skill creates a case and returns a case number, reinforcing an operational behavior that contradicts the documented non-submission boundary. Examples are especially dangerous because implementers often treat them as canonical behavior, which can normalize unsafe automation of regulated banking actions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The document states the skill is "Chinese-first" and that terminology and output templates assume a PRC banking context. While the PRC jurisdiction framing is documented, the language preference is imposed as a default without offering the user a language choice or explicit opt-in, which matches the language/locale policy violation criteria.

Content

No source excerpt is available for this finding.