Back to skill

Security audit

ThinkTank Skill Finder

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed research-skill finder and bundle installer; it persistently installs selected third-party skills, but that matches its stated purpose and is scoped by the catalog and default restricted mode.

Install this only if you want a tool that can add multiple research-related skills to your agent environment. Review the listed bundle before running it, keep restricted mode unless you intentionally want skills that may require API keys or external services, and remember that installed skills persist under ~/.agents/skills and may influence future agent behavior.

Vulnerability Patterns
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • 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
Findings (4)

subprocess module call

Medium
Category
Dangerous Code Execution
Content
return resolved

    probe_cmd = f"where {command}" if os.name == "nt" else f"command -v {command}"
    probe = subprocess.run(
        probe_cmd,
        shell=True,
        text=True,
Confidence
93% confidence
Finding
The command probe uses subprocess.run with shell=True and interpolates the command name directly into a shell string. Although current callers pass fixed values like 'pwsh', 'powershell', 'bash', and 'sh', the helper is generic and unsafe by design; if reused with attacker-controlled input, it enables shell metacharacter injection and arbitrary command execution.

Lp3

Medium
Category
MCP Least Privilege
Confidence
91% confidence
Finding
The skill instructs the agent to read local files and execute shell commands (`clawdhub`, `python`, `curl`, `jq`) but does not declare any permissions or trust boundaries. This creates a capability mismatch where an agent may perform filesystem access and command execution without explicit user-visible authorization, increasing the risk of unintended installs, remote content retrieval, or environment manipulation.

Intent-Code Divergence

Medium
Confidence
84% confidence
Finding
A misleading skill name can cause users or downstream automation to select a tool under false assumptions about its data source or trust properties. Here, 'CNKI Scholar' is documented as not actually integrating with CNKI, which creates a provenance/confusion risk in a skill-finder context where users may rely on exact source coverage.

Intent-Code Divergence

Medium
Confidence
80% confidence
Finding
A name/function mismatch can mislead users into believing a skill performs patent search or scanning when it is only a patentability-analysis template. In a recommendation or installation workflow, this can result in incorrect tool selection, bad research outputs, and misplaced trust in unavailable capabilities.

Static analysis

No suspicious patterns detected.