ClawHub
Back to skill

Security audit

Corporate Banker Digital Employee

Security checks across malware telemetry and agentic risk

Overview

The skill is a banking workflow assistant, but it under-discloses live system access, sensitive data handling, and persistent records that appear throughout the detailed instructions.

Install only in an environment where bank API access, customer-data permissions, audit retention, external data use, and note-saving behavior are explicitly approved and technically scoped. Do not treat this as an offline reference-only skill; require clear user consent before saving notes, creating cases, or using external data sources.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (4)

Description-Behavior Mismatch

High
Confidence
98% confidence
Finding
The skill explicitly claims there is no persistent storage, network access, or credential-like data collection risk, yet the workflow later instructs API calls, external search, file reads, and audit-log creation. This mismatch is dangerous because operators or enforcement layers may trust the safety declaration and enable the skill in contexts where data exfiltration, unexpected system access, or retention should be prohibited.

Intent-Code Divergence

High
Confidence
97% confidence
Finding
Module 1 says it does not perform real-time data queries and relies on preloaded data, but the workflow directs live queries to internal systems and external search sources. That inconsistency can bypass operator expectations, causing the model to access sensitive banking data or external sources under a false assumption that the module is offline/reference-only.

Intent-Code Divergence

Medium
Confidence
88% confidence
Finding
Module 2 states it does not automatically fetch financial data, yet it depends on external APIs or data subscriptions for industry benchmarks and valuation references. This is a real policy-integrity issue because users may believe analysis is limited to uploaded documents while the skill actually reaches outside approved inputs, potentially introducing data leakage, licensing/compliance issues, or unreviewed external influence.

Context-Inappropriate Capability

Medium
Confidence
93% confidence
Finding
The visit-memo module is framed as simple note-taking from user-described information, but it also saves structured notes to bank systems and automatically generates audit logs. In a banking context, silent persistence of visit content can create privacy, retention, and consent risks, especially when notes may contain sensitive customer, operational, or compliance information.

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.