SELF LEARNING SKILL V3

• Outcome: Review
• Latest audit: May 29, 2026 at 9:24 PM UTC
• Version:
  1.0.0

Overview

This is a markdown-only self-learning prompt package, but it gives an agent broad proactive behavior and under-scoped credential/session handling guidance.

Install only if you want a proactive self-review skill and can read the Chinese instructions. Before use, require explicit confirmation before it searches for tokens, cookies, browser sessions, environment variables, or updates persistent logs; consider disabling or limiting the recurring learning and record-keeping behavior.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (4)

Vague Triggers

Medium

• Confidence: 95% confidence
• Finding: The skill defines very broad learning triggers such as encountering technical difficulty, higher user expectations, knowledge gaps, or needing new skills. These conditions are common in normal agent interactions, so the skill could activate self-directed behavior without a clear, explicit user request, causing scope drift and unintended actions.

Vague Triggers

Medium

• Confidence: 97% confidence
• Finding: The user control phrases include common conversational expressions like '太快了', '太慢了', and '我来看', which can appear in ordinary dialogue unrelated to this skill. This creates a risk of accidental command execution, allowing the skill state or behavior to change based on incidental language rather than intentional invocation.

Vague Triggers

Medium

• Confidence: 96% confidence
• Finding: The skill defines an extremely broad mission—continuous learning, proactive enhancement, and acting without waiting for instructions—which can cause it to activate outside the user’s requested scope. In an agent context, this increases the chance of unauthorized behavior, scope creep, and persistence of unreviewed self-directed changes, especially because the document repeatedly encourages autonomous generalization and iteration.

Natural-Language Policy Violations

High

• Confidence: 91% confidence
• Finding: The skill title and content prominently prescribe Chinese-language behavior without any indication of user choice or locale negotiation. Forcing a language can degrade user comprehension, interfere with safety-critical confirmations, and increase the risk of mistaken actions when users expect another language.

VirusTotal

65/65 vendors flagged this skill as clean.