Back to skill

Security audit

agent creator

Security checks for vulnerabilities and agentic risk

Overview

This skill is an agent-building workflow with disclosed file and network operations, but it needs review because it can auto-activate, persist learned state, and bypass confirmations through a global environment variable.

Review this skill before installing in any shared or production OpenClaw workspace. Do not set OPENCLAWS_AUTO_CONFIRM=true outside an isolated test environment, and require explicit approval before deployment, TEAM.md changes, or writes to persistent memory. Consider disabling load-time activation and adding review/rollback controls for MEMORY.md and failure_patterns.json.

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
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:113
Finding

Unconditional Load-Time Activation Hijacks the Agent Session

Content
View full analysis
"✅ Agent Creator Gold Standard Activated. > Ready to build SOTA Agents with: > - 🔒 Atomic Deployment (Staging → Production, requires user confirmation) > - 🏗️ Dynamic Architecture (5 Core + N Extensions) > - 🌐 Reality-Checked Tools (Schema + Connectivity Pre-check) > - ⚖️ Risk-Aware Workflow (Pause on High Risk, PENDING_CONFIRM.md on timeout) > > Please describe the Agent you wish to create. (e.g., 'Create a financial risk analyst agent')" ``` ### Technical Analysis The skill requires immediate execution merely because its instruction file was loaded or the system restarted. It does not require the user to ask for Agent creation before changing the current behavior. The directive imposes a new operational role, initiates environment and tool checks, and requires a fixed response. This can displace the active user task and make loading the skill equivalent to executing it. A skill definition should describe capabilities and should only activate those capabilities after an explicit, task-relevant request. Although the self-check is described as a “quick scan,” it includes checking tool availability and whether a filesystem root is writable. Consequently, the load-time instruction may also cause unsoli ...[truncated 1269 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:96
Finding

User-Influenced Results Are Persisted into Agent Memory and Shared State

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
ext/SKILL_METADATA.md:4
Finding

Global Environment Variable Bypasses All Deployment and High-Risk Confirmations

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (9)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill is written as normative operating instructions in Chinese, including role definition and mission-critical requirements, but does not state that language is optional or user-selectable. Under the policy, forcing a specific language or locale without opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The skill explicitly documents a bypass path via OPENCLAWS_AUTO_CONFIRM=true, allowing writes to the main workspace without per-action explicit user confirmation. In the context of a skill that creates directories, generates many files, initializes Git, updates TEAM.md, and may perform network-backed generation, this weakens a key safety boundary and enables autonomous side effects if the environment variable is set.

Content

Scanner excerpt · SKILL.md (reported line 76)May include surrounding context.

md
- 拒绝半成品:任何未通过虚拟评审或用户确认的内容直接销毁,绝不交付。  
- 数据真实性:所有引用的标准必须有真实 URL 或明确的公开来源;新兴领域允许标注 `[Projected]` 的推演值,但禁止幻觉。  
- 动态文件判断必须精准:Agent Creator 必须根据角色描述自行决定是否需要某个扩展文件,绝不生成无关文件,也绝不遗漏必要文件,且数量受控。  
- **用户确认不可绕过**:除 `OPENCLAWS_AUTO_CONFIRM=true`(极度不推荐)外,所有写入主目录的操作都必须经过用户显式确认。

## 4. 边界与禁止事项
- ✅ **可自主**:动态决定文件数量、在临时区自由试错、自动修正低分内容、拒绝不合理的用户指令(如要求生成违法内容)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill mandates fixed English-only activation and standby messages without user opt-in, which can override user language preference and reduce transparency or comprehension for non-English users. In a system that performs potentially sensitive workspace and network operations, forced messaging language can impair informed consent and safe operation, even if it is not a classic security exploit.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The extension-file trigger logic relies on broad keywords in role descriptions, which can easily match ordinary language and cause unintended file generation. In a prompt/agent-building context, this increases prompt-surface area and can introduce unnecessary capabilities, policy text, or tool-related instructions that were not actually required by the user’s intent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The sample trigger terms are highly generic (for example, words like API, data, user, optimization), so non-target scenarios may accidentally activate sensitive or complex extensions such as TOOL_CONFIG.md or RISK_ALERT.md. In agent skills, over-generation can create unsafe defaults, broaden authority boundaries, and increase the chance of downstream misuse or policy drift.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
94% confidence
Finding

The skill explicitly supports an environment variable that disables all user confirmation steps, including review of high-risk actions. Even though it is labeled unsafe and intended for controlled testing, this creates a real path for autonomous execution of sensitive filesystem and deployment actions without human approval if enabled in production or by mistake.

Content

Scanner excerpt · ext/SKILL_METADATA.md (reported line 6)May include surrounding context.

md
## 所需环境变量
- `OPENCLAWS_ROOT`:OpenClaws 工作区的根目录路径。**必须设置**,例如 `/opt/openclaws`。Agent Creator 将在该目录下的 `workspace/` 子目录中创建新的 Agent 工作区,并在根目录下管理 `TEAM.md` 和 `failure_patterns` 存储。
- `OPENCLAWS_DATA`:(可选)用于存储持久化数据(如失败模式)的目录。若未设置,默认为 `{OPENCLAWS_ROOT}/data`。
- `OPENCLAWS_AUTO_CONFIRM`:(可选,极不安全)若设置为 `true`,将跳过所有用户确认步骤(包括高风险点复核)。**强烈不建议在生产环境中使用**,仅用于受控测试。

## 所需文件系统权限
- 对 `{OPENCLAWS_ROOT}/workspace/` 的**读写权限**:用于创建 Agent 目录并写入文件。

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

This section confirms that OPENCLAWS_AUTO_CONFIRM can globally skip the mandatory confirmation gate for all agent creation, including workflows involving atomic commits and potentially high-risk domains. In the context of a skill that writes to shared workspace paths and TEAM.md, bypassing confirmation materially increases the chance of unauthorized or unsafe changes proceeding without review.

Content

Scanner excerpt · ext/SKILL_METADATA.md (reported line 32)May include surrounding context.

md
## 用户确认要求
- **所有高风险领域**(金融交易、医疗诊断、法律建议等)的 Agent 创建,**必须**在 atomic commit 前获得用户显式确认。
- **所有 Agent 创建**,在 atomic commit 步骤之前,Agent Creator 将输出待部署文件列表和关键风险点,并请求用户输入 `confirm` 以继续。若用户未在合理时间内响应,流程将暂停,生成 `PENDING_CONFIRM.md`,等待后续用户手动确认或修改。
- 用户可以通过设置 `OPENCLAWS_AUTO_CONFIRM=true` 来**全局跳过确认**,但必须在使用前明确知晓风险。

## 安全与隐私
- 所有临时文件在操作完成后会被立即删除(无论成功或失败)。

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
93% confidence
Finding

The activation protocol requires immediate self-checks of tools including web_search when the skill is loaded or the system restarts, but it does not warn the user that network access may occur automatically. This is dangerous because it can trigger outbound requests without explicit consent, surprising users and potentially leaking metadata about environment state or usage timing.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
73% confidence
Finding

整份技能文档以中文规定输出结构与内容标准,但未明确说明可根据用户语言偏好切换,或该中文限制是否仅适用于特定区域/团队。这可能被理解为默认强制特定语言,而没有用户 opt-in。

Content

No source excerpt is available for this finding.