Self Improving 1.2.16

• Outcome: Pass
• Latest audit: Jun 26, 2026 · 11:44 AM UTC
• Version: 2.0.0

Overview

This is a disclosed local memory skill that stores agent corrections and preferences on disk, with privacy considerations but no evidence of hidden or malicious behavior.

Install this only if you want a long-term local memory layer for corrections and preferences. Review ~/self-improving/ periodically, avoid storing secrets or sensitive personal/project data, and use stricter confirmation settings if you want approval before entries are saved.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (11)

Vague Triggers

Medium

• Confidence: 84% confidence
• Finding: The activation criteria are broad enough to trigger this skill during many normal interactions, including routine corrections or self-evaluation after ordinary tasks. In context, that matters because activation causes persistent local storage behavior, so over-broad invocation increases the chance of collecting user preferences or corrections without sufficiently explicit user awareness or consent.

Missing User Warnings

Medium

• Confidence: 95% confidence
• Finding: The skill instructs automatic logging of corrections and preferences to local files, but the description and user-facing setup do not prominently warn that persistence happens automatically. This creates a consent and transparency gap: users may provide feedback assuming it is ephemeral, while the agent retains it across sessions.

Missing User Warnings

Low

• Confidence: 88% confidence
• Finding: The template explicitly instructs creating directories and files under the user's home directory on first activation, which causes persistent filesystem modification without any embedded warning, confirmation step, or scope limitation. In the context of a self-improving agent skill, this is more sensitive because the skill encourages autonomous persistence and recurring state updates, increasing the chance of unexpected writes or privacy concerns.

Missing User Warnings

Medium

• Confidence: 93% confidence
• Finding: The skill describes automatically loading, modifying, and persisting memory across sessions without any explicit consent, notice, or confirmation boundary. That creates a privacy and integrity risk because user-provided content can be silently stored, altered, and later reused or exposed in future interactions.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The weekly maintenance flow automatically moves, archives, and compacts stored data without warning the user about retention changes or data lifecycle transitions. This is dangerous because data may persist longer than expected, move into broader storage surfaces, or become harder for the user to track and delete.

Ssd 3

Medium

• Confidence: 93% confidence
• Finding: These instructions direct the agent to retain user-provided information over time and later surface it through queries like 'What have you learned?' or 'Show my patterns.' Even though storage is local and some categories are excluded, persistent retention of user preferences and corrections can expose sensitive behavioral data, create profiling risks, and increase damage if the local store is accessed by other tools or users.

Ssd 3

Medium

• Confidence: 95% confidence
• Finding: The skill explicitly logs corrections and patterns, writes them to persistent files, and later reuses them across sessions, creating a natural-language memory channel that can leak sensitive user information. Because these entries are intended to be searchable and citable, any secrets, personal preferences, or project details captured there may be surfaced outside their original context.

Ssd 3

Medium

• Confidence: 94% confidence
• Finding: Commands such as showing memory and exporting archives expose a direct retrieval path for accumulated stored content. If memory contains sensitive prior-session data, these commands can disclose it wholesale, making any over-collection or accidental storage materially more dangerous.

Ssd 3

Medium

• Confidence: 95% confidence
• Finding: The skill normalizes persistent storage of corrections, preferences, and reusable lessons across tasks without defining sensitivity limits, consent boundaries, or retention controls. This can lead to inadvertent capture of secrets, personal data, internal project details, or other sensitive natural-language content in long-lived memory files.

Ssd 3

Medium

• Confidence: 96% confidence
• Finding: These AGENTS.md additions encourage broad logging into durable memory locations and treat many kinds of user-provided information as candidates for persistence. Without strict sensitivity boundaries, this creates a cross-task data retention channel that may expose confidential context later or cause the agent to reuse sensitive information inappropriately.

Ssd 3

Medium

• Confidence: 97% confidence
• Finding: The write rules instruct immediate persistence of corrections, preferences, and lessons before the final response, which increases the chance of storing sensitive information reflexively and without review. Because these entries are framed as reusable across domains and projects, the retained data may later influence unrelated tasks and widen exposure.

VirusTotal

64/64 vendors flagged this skill as clean.

Static analysis

No suspicious patterns detected.