Agent自我反思与进化器

• Outcome: Review
• Latest audit: Jul 28, 2026 · 6:09 AM UTC
• Version: 3.0.0

Overview

This is a disclosed local memory and self-improvement skill, but it needs review because it can automatically persist conversation details and change long-term agent guidance.

Review this before installing. Use it only if you want the agent to read daily memory notes, write persistent learning/error files, and potentially update long-term agent guidance. Disable or avoid scheduled runs until you are comfortable with the scope, require human review for MEMORY/SOUL/TOOLS/AGENTS changes, and do not allow tokens, secrets, auth headers, or detailed credential state to be stored in logs.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Rogue Agent: Self-Modification, Session Persistence
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (14)

Vague Triggers

High

• Confidence: 93% confidence
• Finding: The trigger section uses broad everyday phrases like '好', '可以', '对', '能不能', and similar common language, which can cause unintended activation during ordinary conversation. In a skill that writes to persistent memory and logs, accidental activation increases the chance of unnecessary data retention and unwanted file modifications.

Vague Triggers

Medium

• Confidence: 88% confidence
• Finding: The manifest description advertises broad trigger words such as '记录错误', '自我反思', and 'self-improve' without defining activation boundaries. This makes it easier for the skill to engage in contexts where the user did not intend durable memory updates or self-modification behavior.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The skill mandates full-file overwrite via read→write and even acknowledges concurrent-write risk later, but the early instruction is framed as an absolute rule without strong safety guardrails. Because the workflow modifies persistent files and archives entries, users may suffer silent data loss or overwritten changes if multiple sessions operate concurrently.

Missing User Warnings

Medium

• Confidence: 86% confidence
• Finding: The quick-start explicitly instructs users to schedule unattended recurring automation for diary reading and memory updates, but the page provides no warning about what data will be read, what files will be modified, or what summaries may be emitted. In a self-evolution skill, this increases the risk of silent processing of sensitive conversation data and unintended persistent state changes once the cron job is enabled.

Missing User Warnings

Medium

• Confidence: 89% confidence
• Finding: The skill explicitly instructs the agent to write to persistent memory files and later sections direct archival and overwrite operations across memory stores, but it provides no requirement for explicit user confirmation, preview, or dry-run before modifying long-term data. In an agent setting, this can cause unintended persistence changes, silent corruption of memory, or destructive archival behavior triggered by routine execution.

Ssd 3

Medium

• Confidence: 95% confidence
• Finding: The skill instructs the agent to extract information from daily notes and interactions, persist it into multiple long-lived files, and surface summaries back to the user. That creates a real natural-language data leakage risk because sensitive personal, operational, or contextual details can be copied into broader memory stores and later resurfaced in unrelated contexts.

Ssd 3

High

• Confidence: 98% confidence
• Finding: The skill explicitly says to record API and token status in persistent error logs, which can capture sensitive authentication or operational details that should not be stored in plaintext memory files. If those logs are later read, summarized, shared, or archived, they can expose secrets or help an attacker understand the environment and failure modes.

Self-Modification

High

• Category: Rogue Agent

Content

AI Agent 很擅长发现模式，但很难判断哪些模式值得永久记住。Agent 可能会说「我学到这条规则很重要！」——但你怎么知道它不是在幻觉式自信？

mu-self-evolve 把「提议」和「验证」分开。Agent 提议一条规则，然后 Python 脚本去查实际数据：这个错误复现了几次？跨了几个不同任务？提议的方案真的被确认有效吗？只有数据撑得住，规则才能写入永久记忆。

就像一个学生说自己掌握了某个知识点——你不能光听他说，得给他做套真题检验一下。

• Confidence: 80% confidence
• Finding: self-evolve

Self-Modification

High

• Category: Rogue Agent

Content

大多数记忆系统要么永远保留所有东西（直到你手动删除），要么按行数截断（文件太长就删最老的）。两种都不好：永远保留意味着噪声淹没信号；按行截断可能把一条关键规则删了，仅仅因为它比较老。

mu-self-evolve 用**时间衰减公式**——每条记忆的重要性分数会随时间自然下降，就像你忘了几个月前读过的书的细节，但对话中反复被提到的教训却记忆犹新。

被引用到的记忆会重置新鲜度。30 天没人碰的记忆分数减半。但在多个任务中被证明有用的记忆永久豁免——它们已经赢得了自己的位置。

• Confidence: 79% confidence
• Finding: self-evolve

Self-Modification

High

• Category: Rogue Agent

Content

git clone https://github.com/muippt/mu-self-evolve.git

# 2. 初始化（自动检测环境）
cd mu-self-evolve
python3 scripts/env_detect.py --init

# 3. 设置定时进化

• Confidence: 83% confidence
• Finding: self-evolve

Self-Modification

High

• Category: Rogue Agent

Content

---
name: mu-self-evolve
version: "3.0"
description: "AI Agent 持续进化系统：每日经验沉淀+每周错误反思，正向提炼与负向纠偏合一。v3.0 新增 VFM 确定性验算、评分+衰减淘汰、WHERE×WHY 病理归档、主动 Skill 合成。触发词：记录错误、进化系统、自我反思、踩坑记录、self-evolve、self-improve。"
tags: 自我进化,记忆管理,错误记录,持续学习,经验沉淀,自动化,工作流程,反思复盘

• Confidence: 87% confidence
• Finding: self-evolve

Self-Modification

High

• Category: Rogue Agent

Content

---
name: mu-self-evolve
version: "3.0"
description: "AI Agent 持续进化系统：每日经验沉淀+每周错误反思，正向提炼与负向纠偏合一。v3.0 新增 VFM 确定性验算、评分+衰减淘汰、WHERE×WHY 病理归档、主动 Skill 合成。触发词：记录错误、进化系统、自我反思、踩坑记录、self-evolve、self-improve。"
tags: 自我进化,记忆管理,错误记录,持续学习,经验沉淀,自动化,工作流程,反思复盘
visibility: public
license: MIT

• Confidence: 87% confidence
• Finding: self-evolve

Self-Modification

High

• Category: Rogue Agent

Content

无例外。

# mu-self-evolve · 持续进化系统

把经验、错误、纠正、知识缺口沉淀为持久记忆，正向提炼 + 负向纠偏，形成完整进化闭环。

• Confidence: 89% confidence
• Finding: self-evolve

Self-Modification

High

• Category: Rogue Agent

Content

<div class="card">
<div class="card-icon">🔧</div>
<h3><span data-en="Step 2: Initialize" data-zh="步骤2：初始化">步骤2：初始化</span></h3>
<p><code style="background:var(--bg-code);padding:2px 6px;border-radius:4px;font-size:0.82rem">cd mu-self-evolve && python3 scripts/env_detect.py --init</code></p>
<p style="margin-top:8px"><span data-en="Auto-detects OpenClaw, Claude Code, or other environments." data-zh="自动检测 OpenClaw、Claude Code 或其他环境。">自动检测 OpenClaw、Claude Code 或其他环境。</span></p>
</div>
<div class="card">

• Confidence: 81% confidence
• Finding: self-evolve

VirusTotal

VirusTotal findings are pending for this skill version.

Static analysis

No suspicious patterns detected.