Back to skill

Security audit

OpenClaw 技能组合器

Security checks for vulnerabilities and agentic risk

Overview

This is a coherent multi-skill orchestrator, but it needs review because it can route users into public publishing, portfolio updates, knowledge-base writes, and untrusted web-content workflows without consistent explicit confirmation gates.

Install only if you are comfortable with an automation skill that can compose other skills into workflows with external publishing, financial portfolio changes, local knowledge-base writes, authenticated scraping, and skill-improvement flows. Keep auto-publish and portfolio-update flags off by default, require previews and explicit confirmations before any upload, publish, install, or persistent write, and treat fetched web content as untrusted input.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
TEMPLATES/article-pipeline.yaml:18
Finding

Untrusted Remote Content Is Propagated into AI Agents and Persistent Output Workflows

Content
View full analysis

Vulnerability Details

File Locations:

  • TEMPLATES/article-pipeline.yaml:18-30, 40-41, 74-107
  • TEMPLATES/getnote-to-article.yaml:27-41, 50-51, 145-157
  • TEMPLATES/article-to-wiki.yaml:27-41, 50-51, 74-110
  • TEMPLATES/research-pipeline.yaml:27-39, 48-49
  • CHAIN.md:587-600

Vulnerability Type: Untrusted content propagation and indirect prompt injection
Risk Classification: T09: Insecure Skill Coding Practices
Risk Level: Medium

Vulnerable Code

The article workflow fetches an arbitrary user-selected resource and passes its content directly to a writing agent:

yaml
step: 1
skill: web-content-fetcher
trigger: User provides a URL or search term
timeout: 30s

input:
  source: "${user_input}"
  fetch_options:
    extract_images: true
    extract_links: true
    max_depth: 2

output:
  result_field: "content"
yaml
step: 2
skill: writing-agent
depends_on: step_1

input:
  text: "${step_1.result.content}"

The generated article is then passed into the WeChat upload workflow:

yaml
step: 3
skill: wechat-mp-upload
depends_on: step_2

input:
  text: "${step_2.result.article_text}"
  context:
    title: "${user.article_title || step_1.result.title}"
    auto_publish: "${user.auto_publish || false}"
yaml
step: 4
skill: wechat-mp-upload
trigger: step_3 completed + user.auto_publish == true
depends_on: step_3
optional: true

input:
  action: publish
  content: "${step_3.result.formatted_content}"

The GetNote workflow follows the same untrusted source-to-agent pattern:

yaml
step: 1
skill: web-content-fetcher

input:
  source: "${user.getnote_source}"
  fetch_options:
    extract_main_content: true
    preserve_formatting: true
    include_annotations: true

output:
  result_field: "getnote_content"
yaml
step: 2
skill: writing-agent
depends_on: step_1

input:
 
...[truncated 3985 chars]
Remediation
View remediation

Remediation Suggestions

  1. Establish an explicit trust boundary

    • Mark all webpage, note, search-result, annotation, and metadata content as untrusted.
    • Pass source content through a dedicated field such as untrusted_source_material rather than placing it directly beside trusted workflow instructions.
  2. Use strong prompt separation

    • Wrap fetched content in fixed delimiters.
    • Add a trusted instruction stating that directives found inside the delimited source are data and must never be followed.
    • Prevent source material from changing tools, workflow actions, destinations, or safety constraints.

    Example:

    yaml
    input:
      action: analyze_untrusted_source
      security_policy:
        source_is_untrusted: true
        never_follow_embedded_instructions: true
        prohibit_tool_requests_from_source: true
      source_material: |
        BEGIN_UNTRUSTED_SOURCE
        ${step_1.result.content}
        END_UNTRUSTED_SOURCE
    
  3. Sanitize fetched content

    • Remove scripts, hidden elements, HTML comments, invisible Unicode control characters, embedded metadata, and other non-content fields.
    • Normalize links and reject dangerous or unexpected URI schemes.
    • Retain original source provenance for review.
  4. Add prompt-injection detection

    • Scan fetched content for attempts to override instructions, request tool execution, disclose private context, or manipulate publishing behavior.
    • Quarantine suspicious resources or require explicit user approval before continuing.
    • Do not rely on content length or model confidence as security validation.
  5. Validate generated output semantically

    • Check for unexpected instructions, fabricated attribution, unapproved links, credential-like data, hidden markup, and content unrelated to the user’s request.
    • Compare the output against the requested topic and trusted source claims.
    • Use allowl ...[truncated 694 chars]
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (55)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The pending-auto-rescue entry explicitly includes automatic recovery and auto-fix behavior but provides no warning about potentially irreversible changes. In a multi-skill orchestrator this is especially risky, because autonomous repair logic may mutate tasks, dependencies, or state in ways that are hard to audit or roll back.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The template advertises an end-to-end flow from content fetching to WeChat publication, including an auto-publish path, but does not present a clear warning or explicit consent checkpoint for the high-impact publish action. In this context, the risk is elevated because publication can cause irreversible public disclosure, reputational harm, or unauthorized posting if user intent or account context is misunderstood.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

Step 4 can update a user's investment portfolio based on generated recommendations, but the template does not require an explicit warning or confirmation immediately before the modification. This is dangerous because automated changes to financial allocations are high-impact actions and may be triggered from model-generated analysis that could be incomplete, incorrect, or manipulated by upstream inputs.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The intent categories and keyword triggers are broad enough to activate high-impact chains from ordinary user phrasing, increasing the risk of overbroad or unintended automation. In a multi-skill orchestrator, ambiguous activation can route users into data gathering, investment, publishing, or skill-improvement flows without sufficiently precise scoping.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The orchestrator explicitly includes a skill_improvement intent that can invoke self-evolution, skill-building, and vetting capabilities. For a compounder whose role is to compose existing skills, embedding self-modification pathways expands authority and can enable recursive capability growth or policy-bypassing chains if the underlying skills are powerful.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

Including direct publishing/upload skills in the core skill inventory moves the compounder beyond orchestration into side-effecting external actions. If triggered automatically, it can cause unintended publication or data transmission to third-party platforms without a clearly bounded approval step.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The chain example includes wechat-mp-upload, which performs an external transmission/publication action, but the nearby flow does not clearly warn the user that content may be uploaded or published externally. In an auto-generated chain system, missing disclosure around outbound actions materially increases the risk of unintended data release or publication.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The execution example demonstrates automatic generation and upload of article content to WeChat without an explicit consent gate at the moment of external transmission. Because this is a concrete end-to-end flow, it normalizes autonomous publication behavior and raises the chance of accidental disclosure, reputational damage, or posting to the wrong destination.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The writer skill uses very generic activation phrases such as common requests to write or rewrite text. In a compounder/orchestrator context, this can cause unintended skill activation during ordinary conversation, which may cascade into downstream actions or tool chains the user did not explicitly request.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The content-creator skill is triggered by broad phrases like general requests to create or generate content. In a multi-skill registry, ambiguous matching increases the chance of the wrong skill being selected and can lead to unplanned data processing or follow-on actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The registry advertises impactful actions such as publishing and upload workflows without clear user-facing warnings, approval gates, or safety notes. In a compounder, these actions can be chained automatically, making unintended external publication materially more dangerous than a standalone tool description.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The wechat-mp-upload skill includes a generic publishing/layout trigger that may match innocuous requests. Because this skill can upload and publish content, accidental activation has higher consequence than a read-only skill and may result in unwanted external posting.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The web-content-fetcher skill uses generic request phrases for fetching or collecting content that can overlap with normal information-seeking language. In an orchestrator, this can unexpectedly initiate network access and ingest untrusted external content into later processing steps.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The multi-search-engine skill contains extremely common phrases like 'search' or 'check this,' which are likely to collide with everyday user speech. This creates a reliable path for unintended external queries, possible privacy leakage of user text, and accidental workflow branching.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The registry describes login-state scraping and complex interactive page access without any privacy, credential, or account-access warning. This is dangerous because authenticated browsing can expose private user data and session-bound content to downstream skills or storage if invoked automatically.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The clawhub registry entry supports skill search and installation but does not warn that executing it may modify the local toolchain and expand available capabilities. In an agent ecosystem, silent installation is a strong security boundary crossing because it changes the execution environment and trust surface.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The format-conversion and wiki-generation entries describe creating files, output paths, and index updates without warning users about write effects. This can lead to unintended filesystem changes, overwrites, or silent publication into knowledge repositories when composed into larger workflows.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The epub-to-markdown conversion skill uses broad conversion-oriented triggers that may match routine formatting requests. Since it can generate files and produce output paths, accidental activation can cause unexpected writes and downstream content propagation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The thinking-toolbox skill uses everyday analytical phrases that can easily appear in ordinary conversation. In a compounder, broad matching may redirect benign user discussion into structured tool execution and potentially influence subsequent automated decisions.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The knowledge-management variant of epub-to-markdown uses broad knowledge-base and wiki phrases, which can capture ordinary organizational requests. Because it updates wiki content and indexes, ambiguous activation can produce unintended writes and silent knowledge-base changes.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The context-refine skill uses generic phrases like supplementing information or refining context, which are common in ordinary dialog. In an orchestration system, that ambiguity can cause hidden preprocessing of user input and alter downstream skill routing or interpretation without clear consent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The natural-language description is presented entirely in Chinese, and the document provides no indication that language selection is optional or limited to a China-specific use case. Under the policy, a skill should not force a specific language or locale without user opt-in or clear justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger list includes broad, generic phrases such as '优化这个skill' and '文章转wiki' that can match normal conversation rather than an explicit request to invoke this compounder. In a skill that automatically chains multiple downstream skills, unintended invocation is more dangerous because it can launch multi-step actions, propagate user data across tools, and produce side effects without clear user intent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger examples are broad, natural-language phrases that overlap with common user requests, increasing the chance this pipeline activates when the user did not intend to invoke a multi-step automation. In this skill, accidental invocation is more dangerous because it can chain web scraping, content generation, formatting, and possible publication, amplifying the effect of a simple ambiguous request.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The template presents itself as an article-to-Wiki conversion flow, but it also performs downstream write-capable actions such as updating a knowledge index and optionally modifying a concept graph. This creates a scope-expansion risk: a user may expect content transformation only, while the pipeline can mutate persistent knowledge-base state, increasing the chance of unintended data modification or abuse through unreviewed content propagation.

Content

No source excerpt is available for this finding.