Security audit

Handy01 Self Improving Agent

Security checks for vulnerabilities and agentic risk

Overview

This skill is a local conversation-improvement logger with some vague auto-analysis wording, but I found no exfiltration, destructive behavior, privilege escalation, or hidden execution.

Install only if you are comfortable with a skill analyzing conversation quality and storing improvement notes in a local markdown log. Avoid logging secrets, personal data, or proprietary content, and periodically review or delete the improvement log if you use it.

A.I.G

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

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands

Findings (8)

Lp3

Medium

Category: MCP Least Privilege
Confidence: 94% confidence
Finding: The skill advertises behaviors that imply file read/write access (for an improvement log and report generation) but does not declare any explicit tool scope or permissions. That mismatch is risky because it obscures the capability boundary, making it easier for the skill to access or modify local files without clear user/admin review.

Vague Triggers

Medium

Confidence: 91% confidence
Finding: The statement that the skill 'automatically analyzes conversations after each session' is overly broad and does not define trigger conditions, scope, consent, or data boundaries. In practice, this can cause unexpected processing of sensitive conversations and make the skill activate in contexts the user did not intend.

Missing User Warnings

Medium

Confidence: 96% confidence
Finding: The skill describes automatic conversation analysis, learning-log storage, and report generation without warning users that their conversation content may be retained on disk. This is dangerous because logs and reports can capture sensitive prompts, personal data, credentials, or proprietary information, especially when default paths are local markdown files.

Description-Behavior Mismatch

Medium

Confidence: 95% confidence
Finding: The manifest describes a self-improving system that identifies improvement opportunities and continuously optimizes response strategies. In practice, the implementation performs simple keyword and length-based analysis, writes notes to a markdown log, and generates reports/suggestions, but it never updates any response policy, model behavior, or strategy configuration.

Missing User Warnings

Medium

Confidence: 89% confidence
Finding: The skill persists conversation-derived improvement insights to a workspace file without any notice, consent flow, retention policy, or data minimization. If users or operators pass sensitive conversation content or identifiers into insights, this creates a quiet privacy leak and expands the local attack surface through accumulated plaintext history.

Vague Triggers

Medium

Confidence: 93% confidence
Finding: The skill description is extremely broad ('Self-improving agent system for continuous optimization') and does not define what inputs, actions, or boundaries apply. For a self-improving or optimizing agent, vague scope is risky because it can justify expansive behavior, autonomous strategy changes, or interaction with other capabilities without clear guardrails, making misuse and unsafe activation more likely.

Intent-Code Divergence

Low

Confidence: 90% confidence
Finding: The module docstring says 'Continuous improvement system' and the class docstring says it 'improves agent performance over time,' which implies active improvement behavior. The code itself only analyzes conversations, appends notes to a log, and suggests possible SOUL.md updates, so the documentation materially overstates what the implementation does.

Natural-Language Policy Violations

Low

Confidence: 83% confidence
Finding: The heuristic feedback analysis is explicitly limited to specific Chinese and English keywords, which bakes in language assumptions without offering language choice or documenting that the skill is intended only for those locales. This can create inconsistent behavior for users in other languages and constitutes a locale policy concern.

Static analysis

No suspicious patterns detected.