ClawHub
Back to skill

Security audit

Multi-Skill-Eval | 集成化技能评估系统

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a real skill-evaluation toolkit, but it needs Review because some broad benchmark/rewrite workflows and generated HTML outputs are not safely bounded.

Install only if you are comfortable running evaluator scripts on skill directories you choose. Use benchmark mode only for trusted or sandboxed target skills, review any generated skill rewrite before making it active, and avoid opening generated leaderboard HTML from untrusted skill cards until the HTML escaping issue is fixed.

SkillSpector

Vulnerability Patterns
  • Rogue AgentSelf-Modification, Session Persistence
  • 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 (1)

Self-Modification

High
Category
Rogue Agent
Content
- Overhead Routing: Add quick-mode vs full-framework routing
   - Assertion-Aligned Rewrite: Rewrite to pass specific failed assertions

4. **Rewrite SKILL.md** with selected strategy:
   - Default: Remove > Add (delete 60-80% first, then add behavioral mandates)
   - Add specific, enforceable conventions
   - Remove redundant content model already handles
Confidence
96% confidence
Finding
The skill explicitly proposes rewriting SKILL.md as part of a self-evolution workflow. Self-modification is dangerous because it allows an agent to alter its own future behavior and trust boundaries, potentially persisting prompt injection, weakening safeguards, or introducing malicious instructions that survive beyond the current session.

Static analysis

No suspicious patterns detected.