ClawHub
Back to skill

Security audit

生物医药专利检索与组合研究（Biopharma Patent Landscape Research）

Security checks for vulnerabilities and agentic risk

Overview

The skill’s patent-search purpose is coherent, but a legacy reference contains conflicting Excel instructions that could unexpectedly change a user’s workbook structure.

Install only if you are comfortable with an agent editing Excel workbooks for patent reporting. Use strict-template mode for user-provided templates, explicitly tell the agent not to load or apply legacy Step 11 formatting, and keep a backup of any workbook before use.

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 (2)

Description-Behavior Mismatch

Medium
Confidence
94% confidence
Finding
The document defines strict invariants earlier in the workflow: do not add columns, do not modify non-target columns, and preserve fixed Sheet2 structure. Step 11 later instructs the agent to insert a serial-number column and add formatting-driven structural changes, which can silently shift column positions and corrupt downstream logic or user data because many later steps depend on header/column stability.

Intent-Code Divergence

High
Confidence
97% confidence
Finding
Step 11 directly contradicts earlier guardrails by modifying Sheet2 globally, inserting a new first column, and restyling all cells after prior steps rely on fixed headers and non-target-column preservation. This creates a dangerous instruction conflict where an agent following the later section may break data integrity, invalidate earlier header lookups, or overwrite formatting/hyperlink semantics in ways that are hard to detect.

Static analysis

No suspicious patterns detected.