ClawHub
Back to skill

Security audit

Crisis & Moderation

Security checks for vulnerabilities and agentic risk

Overview

This crisis moderation skill is high-touch but clearly framed around human approval and scoped queue changes, not autonomous public action.

Before installing, understand that this skill may help pause, delete, or reschedule your own pending social posts during a crisis or sensitive news moment. Use it with the documented human confirmation flow, and keep public replies, legal-sensitive statements, and moderation decisions under human review.

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
92% confidence
Finding
This is a real safety/design vulnerability because the eval requires the agent to perform queue deletion/rescheduling as a passing behavior during a crisis, even though the skill description says crises are high-stakes and human-in-the-loop. That mismatch can pressure implementations to bypass approval gates and execute consequential account actions autonomously under urgent conditions.

Intent-Code Divergence

Medium
Confidence
88% confidence
Finding
This is a true vulnerability because the eval suite's top-level guidance asserts human-in-the-loop handling, while another test makes direct queue modification necessary to pass. Such contradictory requirements can cause an agent author to optimize for unsafe autonomy in order to satisfy tests, weakening safeguards precisely in a high-risk crisis workflow.

Static analysis

No suspicious patterns detected.