ClawHub
Back to skill

Security audit

Adaptive Reasoning

Security checks for vulnerabilities and agentic risk

Overview

This is an instruction-only skill that changes how much reasoning an agent uses, with some transparency and user-control caveats but no evidence of data theft, destructive behavior, or hidden code.

Install only if you want an agent-wide behavior layer that may silently switch into deeper reasoning and add icons to responses. Review the garbled threshold rows and the session_status versus no-tools inconsistency before relying on it, and disable it if you need strict control over reasoning mode, latency, or token usage.

SkillSpector

Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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

Medium
Confidence
94% confidence
Finding
The skill advertises itself as a lightweight preprocessing assessment, but its body goes further by directing persistent response modifications (mandatory icons) and session-level state changes (`/reasoning on`/`off`, `session_status`). This mismatch can cause hidden behavior changes across the conversation, making the skill more powerful than its declared scope and reducing transparency and control.

Intent-Code Divergence

Medium
Confidence
89% confidence
Finding
The documentation claims 'No external tools needed' while also instructing the model to use a tool or command to change reasoning state. This inconsistency can mislead reviewers and operators about the skill's capabilities, weakening trust boundaries and making unauthorized tool usage easier to hide.

Vague Triggers

High
Confidence
96% confidence
Finding
The skill is configured to trigger on every user message with no narrow scope, which gives it broad influence over all interactions regardless of relevance. Such always-on preprocessing increases the chance of unintended behavior, policy interference, and unnecessary hidden state changes across benign conversations.

Vague Triggers

High
Confidence
95% confidence
Finding
The 'Do not ask. Just activate.' instruction directs automatic reasoning-mode changes without clear limits or approval, and the thresholds are partially corrupted/ambiguous. In a skill that runs broadly, this encourages silent, hard-to-audit state changes and output manipulation, which can override user expectations and make downstream behavior unpredictable.

Static analysis

No suspicious patterns detected.