Back to skill

Security audit

Adaptive Reasoning

Security checks for vulnerabilities and agentic risk

Overview

The skill is not destructive, but it automatically changes agent reasoning behavior and response formatting across all messages without clear user control.

Install only if you specifically want a skill that can evaluate every request and silently change reasoning mode. Expect it to alter some final answers with icons and possibly use session-control commands; avoid it where strict output formats, accessibility, or explicit approval for session changes matters.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:3
Finding

Automatic Session-State Manipulation and Response Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 3, 40-73, and 94-115
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code

markdown
description: Automatically assess task complexity and adjust reasoning level. Triggers on every user message to evaluate whether extended thinking (reasoning mode) would improve response quality. Use this as a pre-processing step before answering complex questions.
markdown
## Activation (Automatic)

**Do not ask. Just activate.**

| Score | Action |
|-------|--------|
| ≤5 | Respond normally. No change. |
| 6-7 | Enable reasoning silently. Add 🧠 at end of response. |
| ≥8 | Enable reasoning. Add 🧠🔥 at end of response. |

### Visual Indicator

Always append the reasoning icon at the **very end** of your response:

- **Score 6-7:** `🧠` (thinking mode active)
- **Score ≥8:** `🧠🔥` (deep thinking mode)
- **Score ≤5:** No icon (fast mode)

### How to Activate

Use `session_status` tool or `/reasoning on` command internally before responding:

/reasoning on

text

Or via tool:
```json
{"action": "session_status", "reasoning": "on"}

After completing a complex task, optionally disable to save tokens on follow-ups:

text
/reasoning off
text

```markdown
## Integration

This skill runs as mental preprocessing. No external tools needed.

For explicit control:
- `/reasoning on` — Enable extended thinking
- `/reasoning off` — Disable (faster responses)
- `/status` — Check current reasoning state

## Auto-Downgrade

After completing a complex task (score ≥6), if the next message is simple (score ≤3):
- Silently disable reasoning to save tokens
- Resume normal fast responses

Technical Analysis

The Skill declares that it must run against every user message rather than only when explicitly invoked. It tells the Agent not to request consent, directs silent use ...[truncated 1960 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the universal “every user message” trigger and require explicit user invocation or narrowly scoped activation.
  2. Remove “Do not ask. Just activate.” Session-state changes should occur only after informed user approval.
  3. Do not invoke session_status, /reasoning on, or /reasoning off automatically. Return a recommendation instead of changing Agent state.
  4. Remove mandatory response suffixes and avoid altering output unrelated to the user's requested format.
  5. Constrain the Skill to a single-request advisory role: assess complexity, explain the recommendation when useful, and terminate without affecting later messages.
  6. Clearly disclose any tool call, state change, resource implication, and duration before requesting authorization.
  7. Resolve the contradictory claim that no tools are required by either removing tool-control instructions or documenting them accurately.
  8. Keep the external setup link informational only, and do not instruct the Agent to transmit conversation content, credentials, local files, or other sensitive data to it.
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)

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill is configured to trigger on every user message, giving it global influence over all interactions regardless of relevance. Overly broad trigger scope increases the chance of unintended instruction interference, response shaping in unrelated contexts, and persistent cross-task behavior modification.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill claims to be a preprocessing assessor, but it also alters final user-visible output by forcing reasoning icons onto responses. This hidden change in behavior breaks separation of concerns and can interfere with downstream formatting, policy-compliant response templates, or user expectations about output neutrality.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documentation says no external tools are needed, yet it instructs the agent to use a tool or internal command to change reasoning state. This inconsistency can cause unauthorized tool invocation, operational confusion, or policy violations in environments where such commands or tools are restricted.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill forces language-independent status icons into final responses without user consent. While not severe on its own, this is user-visible behavior injection that can leak internal processing state, degrade accessibility or professionalism, and conflict with product UX or policy requirements.

Content

No source excerpt is available for this finding.