Security audit

Adaptive Reasoning

Security checks for vulnerabilities and agentic risk

Overview

This is an instruction-only skill that changes how the agent chooses reasoning depth, with no code, secrets, network use, or file access, but it may silently change response style and session reasoning behavior.

Install only if you want the assistant to automatically manage reasoning depth across the conversation. Expect possible extra latency or token use on complex prompts and visible reasoning icons at the end of some responses; modify or disable those instructions if you prefer explicit consent or clean output formatting.

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: 93% confidence
Finding: The skill claims to be a preprocessing-only complexity assessment, but it also directs the agent to change session reasoning state and alter user-visible output by appending icons. This hidden coupling between internal assessment and external behavior can override system/user preferences, create unapproved side effects on every message, and make the assistant behave inconsistently with its declared purpose.

Intent-Code Divergence

Medium

Confidence: 90% confidence
Finding: The file states that no external tools are needed, yet instructs use of a tool or command to change session state. This mismatch is dangerous because it hides operational behavior from reviewers and users, increasing the chance of unauthorized state changes and making the skill harder to reason about safely.

Vague Triggers

High

Confidence: 96% confidence
Finding: An automatic trigger on every user message gives the skill broad authority over the full conversation, including benign or low-risk prompts that do not need intervention. In context, this is more dangerous because the skill also attempts to alter reasoning mode and output formatting, so a global trigger creates persistent, hard-to-audit behavioral manipulation across the session.

Natural-Language Policy Violations

Medium

Confidence: 91% confidence
Finding: Forcing a visual indicator into every qualifying response without user opt-in modifies assistant output in a way the user did not request. This can leak internal orchestration state, create confusing or policy-inconsistent responses, and serve as a covert signaling channel for hidden process state.

Static analysis

No suspicious patterns detected.