Back to skill

Security audit

evolving skill creator

Security checks for vulnerabilities and agentic risk

Overview

The skill has a coherent self-improving skill-builder purpose, but it teaches broad permissions and unattended scheduled agent runs that need careful review before installation.

Install only if you want a skill that can create and evolve other skills. Before using it, keep permissions minimal, do not use the full-access template or permission-skip mode, avoid unattended cron/launchd jobs until commands and paths are reviewed, and inspect every generated settings.local.json and child-agent upgrade plan.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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
Findings (2)

T06 · System Persistence

Error
Location
memory/knowledge/agent-design/auto-tick-mechanism.md:18
Finding

Persistent Unattended Agent Execution with Broad Tool Authorization

Content
View full analysis
> /tmp/evo-skill-tick.log 2>&1 0 8 * * 1 cd /path/to/project && claude -p "/evo-skill-creator scan" --print >> /tmp/evo-skill-scan.log 2>&1 0 17 * * 5 cd /path/to/project && opencode run -m opencode/mimo-v2-pro-free "/evo-skill-creator review" >> /tmp/evo-skill-review.log 2>&1 0 * * * * cd /path/to/project && opencode run -m opencode/nemotron-3-super-free "/evo-skill-creator learn" >> /tmp/evo-skill-learn.log 2>&1 0 17 * * 5 cd /path/to/project && claude -p "/evo-skill-creator review" --print >> /tmp/evo-skill-review.log 2>&1 ``` The same document also presents `--dangerously-skip-permissions` as an available authorization option. ### Technical Analysis The document supplies directly actionable instructions for installing operating-system-level scheduled tasks. These tasks survive the original Skill invocation and automatically relaunch an AI agent in a non-interactive context. The first scheduled command preauthorizes a broad collection of capabilities: - Arbitrary shell interaction through `Bash` - File creation and modification through `Write` and `Edit` - Filesystem discovery through `Glob` and `Grep` - External network research through `WebSearch` and `WebFetch` This is especially risky when combined with the behavior in `SKILL.md:47-48`, where high-priority backlog items may be executed without first asking the user. Although the persistence document asks the owner to confirm whether scheduling should be configured, the recommended authorization scope exceeds the minimum privileges needed for simple periodic status checks or resear ...[truncated 1677 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
references/settings.local.json.template.md:15
Finding

Generated Permission Templates Permit Self-Modification and Unrestricted Home-Directory Access

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (88)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The natural-language activation phrases are extremely broad and overlap with ordinary conversation, making accidental invocation likely. Because this skill can read private memory, write logs/reports, and coordinate other agents, misfires can trigger sensitive data access or state changes without the user realizing the skill was activated.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 97)May include surrounding context.

md
- [ ] `SKILL.md` — 符合标准章节模板

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Mentioning --dangerously-skip-permissions in an automation guide without a strong, immediate warning is dangerous because it normalizes bypassing the last interactive safeguard for an agent that may execute on a schedule. If adopted, the agent could perform unrestricted actions repeatedly and silently, greatly amplifying the blast radius of prompt mistakes, compromised inputs, or malicious skill evolution.

Content

No source excerpt is available for this finding.

External Model or Provider Selection

High
Category
Excessive Agency
Confidence
90% confidence
Finding

Skill selects an external model or provider that may use a different account or billing plan than the operator expects. Undisclosed model switches can cause unexpected cost or quota consumption.

Content

Scanner excerpt · memory/knowledge/agent-design/auto-tick-mechanism.md (reported line 51)May include surrounding context.

tick.sh — 一键执行所有定期任务

cd /path/to/project echo "[$(date)] === tick start ===" >> tick.log claude -p "/evo-skill-creator go" --print >> tick.log 2>&1 echo "[$(date)] === tick end ===" >> tick.log

text

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

The template instructs users to place a settings file under the agent configuration hierarchy in ~/.claude/skills/.../.claude/settings.local.json, normalizing direct manipulation of agent configuration storage. While configuration files must live somewhere, referencing and operationalizing this path is security-relevant because it enables persistent capability changes for future agent runs.

Content

Scanner excerpt · references/settings.local.json.template.md (reported line 3)May include surrounding context.

md
# settings.local.json 权限配置模板

> 创建自我进化智能体时,将此模板复制到 `~/.claude/skills/<skill-name>/.claude/settings.local.json`
> 并替换 `{{SKILL_NAME}}` 为实际技能名称

## 基础版(自我进化智能体必备)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The markdown presents broad filesystem and shell permissions as a normal configuration option without clearly warning that they enable access to most user data and arbitrary command execution. Users copying this template may unknowingly authorize capabilities that make any downstream prompt injection, self-modification bug, or malicious skill behavior far more damaging.

Content

No source excerpt is available for this finding.

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

The usage instructions explicitly direct saving permission settings into the agent's configuration directory, which creates persistent authorization state for the skill. In the context of a self-evolving agent creator, persistent config changes are more dangerous because they can institutionalize overly broad access across iterations.

Content

Scanner excerpt · references/settings.local.json.template.md (reported line 84)May include surrounding context.

md
1. 创建新角色时,根据角色需求选择合适的模板版本
2. 将 `{{SKILL_NAME}}` 替换为实际的技能名称
3. 将文件保存到 `~/.claude/skills/<skill-name>/.claude/settings.local.json`
4. 确保已创建 `.claude` 目录:`mkdir -p ~/.claude/skills/<skill-name>/.claude`

## 权限字段说明

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

该文件的名称、描述、命令示例和交互对象均以中文固定表述,且未说明可根据用户偏好切换语言,也未提供语言选择或 opt-in 机制。按照语言/区域政策要求,这种默认强制单一语言的说明可构成自然语言策略违规。

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill instructs itself to read private data and persist logs/reports, but it does not clearly warn the user at activation time that private memory may be accessed and outputs may be written. This undermines informed consent and increases the chance of silent collection, retention, or propagation of sensitive information.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The wake-up flow mandates reading memory/private/README.md on every invocation, regardless of whether the current command actually needs private context. That creates unnecessary exposure of sensitive data and violates least-privilege principles, especially for simple commands like status or help where private data access is not required.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill simultaneously requires generated SKILL.md files to embed absolute paths while elsewhere forbidding hardcoded personal absolute paths and preferring relative paths in the registry. This ambiguity can lead implementers to store sensitive filesystem locations in distributable artifacts, leaking usernames, home-directory layouts, or other environment details.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The review workflows authorize this skill to inspect, invoke status on, and plan upgrades for other child agents, effectively giving it cross-agent orchestration and visibility. If those agents hold separate context or permissions, this creates an over-broad control plane that can expose data across agent boundaries or trigger unintended actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The document recommends detailed execution tracing, context tracking, and memory/file access monitoring, but does not mention redaction, minimization, retention limits, or access controls. In an agent skill that processes prompts, memory, tool inputs, and user data, this can lead to systematic capture of secrets, personal data, and sensitive business context into logs, increasing breach and secondary-use risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The document content is written in Chinese throughout, including metadata and all substantive guidance, with no indication that the language choice is optional or tied to a region-specific requirement. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The file is written entirely in Chinese, including the title and all instructional content, with no indication that language selection is optional or that the skill is intentionally limited to a Chinese-speaking context. Under the language/locale policy, this can be a natural-language policy violation because it imposes a specific language without user opt-in.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The document gives concrete instructions for unattended scheduled execution of an agent with pre-authorized file/system-capable tools, but does not prominently warn that these runs can modify files, invoke shell commands, fetch remote content, or accumulate unintended side effects without real-time human review. In the context of an 'evolving' self-updating skill, automation materially increases the chance of silent drift, destructive edits, or abuse if prompts, hooks, or project state become unsafe.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
91% confidence
Finding

The cron-based persistence mechanism enables recurring execution of an agent with privileged tools, which creates a genuine persistence and autonomy risk even though it is documented as a feature. In this skill's context, persistence is more dangerous because the agent is explicitly intended to self-learn, scan, plan, and review over time, increasing opportunities for unattended harmful behavior or stealthy repeated changes.

Content

Scanner excerpt · memory/knowledge/agent-design/auto-tick-mechanism.md (reported line 19)May include surrounding context.

bash
# 编辑 crontab
crontab -e

# Claude:每天早上 9:00 自动执行 go 命令
0 9 * * * cd /path/to/project && claude -p "/evo-skill-creator go" --allowedTools "Bash,Read,Write,Edit,Glob,Grep,WebSearch,WebFetch" --print >> /tmp/evo-skill-tick.log 2>&1

Session Persistence

Medium
Category
Rogue Agent
Confidence
89% confidence
Finding

Using launchd to auto-start recurring agent tasks after login creates durable persistence similar to cron, with automatic reactivation after reboot/login. This makes the skill more dangerous because a misconfigured or compromised automation loop can continue operating without fresh user intent, especially when paired with tool permissions and self-directed role behavior.

Content

Scanner excerpt · memory/knowledge/agent-design/auto-tick-mechanism.md (reported line 40)May include surrounding context.

md
**优点**:持久、可靠、OS 重启后自动恢复
**注意**:需要 `--allowedTools` 或 `--dangerously-skip-permissions` 授权工具使用

### 方案二:launchd plist(macOS 原生,比 cron 更推荐)

创建 `~/Library/LaunchAgents/com.evo-skill-creator.tick.plist`,实现登录后自动启动定期任务。

Session Persistence

Medium
Category
Rogue Agent
Confidence
86% confidence
Finding

This finding refers to the same launchd persistence guidance and is substantively the same risk: durable, automatic re-execution of an agent outside immediate user supervision. While the line overlap appears duplicative, the underlying issue is still real because persistence increases the impact of any unsafe prompt, tool access, or self-modification workflow.

Content

Scanner excerpt · memory/knowledge/agent-design/auto-tick-mechanism.md (reported line 42)May include surrounding context.

md
### 方案二:launchd plist(macOS 原生,比 cron 更推荐)

创建 `~/Library/LaunchAgents/com.evo-skill-creator.tick.plist`,实现登录后自动启动定期任务。

### 方案三:Shell 脚本 + 手动/半自动触发

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

This markdown skill file is written entirely in Chinese and presents its framework content as the default interaction format, but it does not state that language choice is optional or limited to a China-specific use case. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · memory/knowledge/agent-design/constrained-self-evolution-framework-2026.md (reported line 21)May include surrounding context.

md
## 理论基础:Impossible Trinity(不可能三角)

来源:[The Impossible Trinity of Agentic AI](https://promptedllc.com/research/the-impossible-trinity-of-agentic-ai) + [Anthropic Safety is Always Vanishing](https://api.emergentmind.com/papers/2602.09877)

### 三角定理

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · memory/knowledge/agent-design/constrained-self-evolution-framework-2026.md (reported line 451)May include surrounding context.

md
## 理论基础:Impossible Trinity(不可能三角)

来源:[The Impossible Trinity of Agentic AI](https://promptedllc.com/research/the-impossible-trinity-of-agentic-ai) + [Anthropic Safety is Always Vanishing](https://api.emergentmind.com/papers/2602.09877)

### 三角定理

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This markdown file contains natural-language instructions and guidance exclusively in Chinese, and it does not indicate that the user may choose another language or that the content is intentionally region-specific. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This markdown file contains user-facing instructional content exclusively in Chinese, and there is no indication that users may opt into another language or that the skill is intended only for a Chinese-language context. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.