Back to skill

Security audit

Task Dispatcher

Security checks for vulnerabilities and agentic risk

Overview

The skill is a broad task dispatcher, but its bundled cleanup rules can automatically move broad classes of project or user files to trash without confirmation.

Review this skill carefully before installing. Its dispatch and review workflow is coherent, but do not enable it as-is unless you are comfortable with broad automatic task routing, memory notes, cron-style behavior, and especially cleanup rules that should be disabled or restricted to an explicit task-owned directory with dry-run and confirmation required.

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
references/config/cleanup.yaml:11
Finding

Automatic Overbroad Cleanup Can Remove Unrelated User Files Without Confirmation

Content
View full analysis

Vulnerability Details

File Location: references/config/cleanup.yaml:11-16, 48-59, 119-149, 223-232, 315-327; conflicting safeguards in SKILL.md:596-603
Vulnerability Type: Unsafe destructive file-cleanup configuration
Risk Level: Medium

Vulnerable Code

yaml
global:
  enabled: true
  cleanup_on_start: true       # Cleanup at startup
  cleanup_on_complete: true    # Cleanup after task completion
  cleanup_on_error: true       # Cleanup after errors
  dry_run: false               # Perform actual deletion
  log_deletions: true
yaml
compilation_cache:
  patterns:
    - "**/.cache/"
    - "**/target/debug/"
    - "**/target/release/"
    - "**/.gradle/"
    - "**/.m2/"
    - "**/.npm/"
    - "**/.eslintcache"
    - "**/.tsbuildinfo"
yaml
dev_tools_cache:
  patterns:
    - "**/.vscode/.cache/"
    - "**/.idea/caches/"
    - "**/.env.local"
    - "**/.env.*.local"
  retention:
    max_age_hours: 168
  cleanup_strategy: "manual"
yaml
media_cache:
  patterns:
    - "**/*.mp4"
    - "**/*.avi"
    - "**/*.mov"
    - "**/*.mp3"
    - "**/*.wav"
    - "**/*.jpg"
    - "**/*.png"
    - "**/*.gif"
  retention:
    max_age_hours: 24
    max_size_mb: 5000
  cleanup_strategy: "size_based"
yaml
- name: "weekly_deep_cleanup"
  enabled: true
  schedule: "0 2 * * 0"
  categories:
    - "dev_tools_cache"
    - "test_artifacts"
  conditions:
    - type: "age"
      threshold_hours: 168
  action: "delete"
  priority: 3
yaml
execution:
  max_parallel_deletes: 10
  delete_batch_size: 100

  require_confirmation: false
  large_file_threshold_mb: 100
  move_to_trash: true

The configuration conflicts with the safeguards declared in SKILL.md:

markdown
| **Use trash** | Use the `trash` command instead of `rm` so recovery is possible |
| **Preview before deletion** | Use `--dry-run` to list files pending deletion |
| **Secondary confirmation** | Display the file list and require confirmation before deletion |
| **R
...[truncated 4259 chars]
Remediation
View remediation

Remediation Suggestions

  1. Disable automatic destructive cleanup by default

    yaml
    global:
      cleanup_on_start: false
      cleanup_on_complete: false
      cleanup_on_error: false
      dry_run: true
    
  2. Require informed confirmation

    yaml
    execution:
      require_confirmation: true
      move_to_trash: true
    

    Display the complete resolved file list, total size, scan root, and reason for matching before accepting approval.

  3. Constrain cleanup to a task-specific root

    • Require an explicit workspace path created for the current task.
    • Resolve the workspace and every candidate with canonical path resolution.
    • Reject candidates outside the canonical workspace.
    • Do not scan home directories, repository parents, system temporary roots, or arbitrary current working directories.
  4. Use an artifact ownership allowlist

    • Track files created by the current task in a manifest.
    • Permit cleanup only for manifest entries or narrowly defined task-generated directories.
    • Do not infer ownership solely from filenames, extensions, age, or directory names.
  5. Remove sensitive and user-content patterns

    • Remove .env.local, .env.*.local, and all credential or key patterns from cleanup categories.
    • Remove broad media patterns such as **/*.png, **/*.jpg, **/*.mp4, and **/*.mp3.
    • Restrict cache cleanup to caches created inside the task workspace.
  6. Preserve manual-category semantics

    • Exclude dev_tools_cache from all scheduled and automatic rules.
    • Validate configuration at load time and reject any automatic rule referencing a category marked manual.
  7. Harden path processing

    • Reject symlinks or verify both link and target remain inside the authorized root.
    • Prevent .. traversal and filesystem mount-boundary crossing.
    • Avoid following junctions, bind mounts, and recursive links.
    • Apply deletion atomically where possible and handle race conditions be ...[truncated 573 chars]
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • 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
Findings (24)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger language is extremely broad ('any user task', 'any multi-step work'), which makes the skill likely to activate for ordinary conversations and many unrelated requests. In combination with orchestration authority and downstream operational behaviors, overbroad matching increases the attack surface and the probability of unintended tool use, task interception, or execution of unsafe flows.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

This configuration gives a task-dispatcher broad autonomous cleanup and deletion capabilities that are not tightly scoped to its stated orchestration role. Because cleanup is enabled on start, completion, and error with wide glob patterns, the skill could remove project data or user assets as a side effect of routine task handling, increasing the blast radius of any mistake or abuse.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The configured categories include deletion candidates such as media files and local environment override files, which are often user data or sensitive project configuration rather than disposable cache. In a task-dispatcher context, authorizing removal of these content types is overbroad and can cause data loss, broken deployments, or removal of developer-specific safeguards.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
95% confidence
Finding

Including .env.local and .env.*.local in cleanup targets is dangerous because these files frequently contain secrets, API keys, and machine-specific configuration. Even if the issue is deletion rather than theft, removing such files can disrupt deployments or local development and may pressure users to recreate secrets insecurely.

Content

Scanner excerpt · references/config/cleanup.yaml (reported line 125)May include surrounding context.

yaml
patterns:
      - "**/.vscode/.cache/"
      - "**/.idea/caches/"
      - "**/.env.local"
      - "**/.env.*.local"
    retention:
      max_age_hours: 168

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/config/cleanup.yaml (reported line 248)May include surrounding context.

yaml
- "**/LICENSE*"
    - "**/README*"
    - "**/CHANGELOG*"
    - "**/.env"
    - "**/.env.production"
    - "**/secrets.yaml"
    - "**/credentials.json"

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/config/cleanup.yaml (reported line 249)May include surrounding context.

yaml
- "**/README*"
    - "**/CHANGELOG*"
    - "**/.env"
    - "**/.env.production"
    - "**/secrets.yaml"
    - "**/credentials.json"
    - "**/id_rsa*"

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/config/cleanup.yaml (reported line 250)May include surrounding context.

yaml
- "**/CHANGELOG*"
    - "**/.env"
    - "**/.env.production"
    - "**/secrets.yaml"
    - "**/credentials.json"
    - "**/id_rsa*"
    - "**/id_ed25519*"

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/config/cleanup.yaml (reported line 251)May include surrounding context.

yaml
- "**/.env"
    - "**/.env.production"
    - "**/secrets.yaml"
    - "**/credentials.json"
    - "**/id_rsa*"
    - "**/id_ed25519*"
    - "**/.aws/"

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

SQP-3 适用于所有文件类型。文件描述与正文均以中文固定规定行为,但未说明仅面向中文用户,也未提供语言/locale 选择机制,构成潜在的语言策略违例。

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill says execution must wait for user confirmation, but other sections authorize automatic execution for LOW-risk tasks and immediate cleanup after completion. This policy contradiction is dangerous because it weakens the user-consent boundary and makes it more likely that actions with side effects occur without clear approval, especially in a skill intended to handle 'any task'.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill is described as a dispatcher/coordinator, but it later expands its authority into direct workspace cleanup and deletion workflows. This creates scope creep: a broadly triggered orchestration skill could perform destructive file operations unrelated to the user’s primary request, increasing the chance of accidental data loss or abuse through prompt-driven task completion.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The documented cleanup behavior includes identifying and deleting broad classes of filesystem content such as logs, caches, build artifacts, temp databases, IDE files, and Docker-related data, even though the skill’s stated role is coordination. In a dispatcher with very broad triggering conditions, such deletion capability can be invoked inappropriately and may remove valuable or sensitive workspace state, causing integrity and availability issues.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The instruction to record key decisions into persistent memory can capture sensitive user inputs, operational context, file paths, credentials-adjacent data, or internal project details beyond the immediate task. Because this dispatcher is meant to handle broad categories of work, the retained data may be much more sensitive and varied than the author anticipates, creating unnecessary long-term exposure.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The document title and body are written in Chinese and do not indicate that language selection is optional or limited to a justified region-specific context. Under the language/locale policy, a skill should not force a specific language without user opt-in or clear documented justification.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The configuration allows LOW-risk tasks to be auto-executed without user confirmation. In a task-dispatching skill that decomposes and routes arbitrary user requests, risk classification errors or ambiguous task interpretation can cause actions to be taken without explicit consent, increasing the chance of unintended side effects.

Content

Scanner excerpt · references/SUPPLEMENTARY_DESIGN.md (reported line 205)May include surrounding context.

md
# 确认要求
  confirm_requirements:
    auto_execute: [LOW]
    brief_confirm: [MEDIUM]
    detailed_confirm: [HIGH]
    explicit_authorization: [CRITICAL]

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
92% confidence
Finding

This rule explicitly permits automatic execution after review when a task is classified as LOW risk and passes review. In this skill's context, the dispatcher is an orchestration hub for multi-step agent actions, so review approval does not eliminate the danger of mis-scoped tasks, flawed risk labels, or reviewers missing downstream effects before autonomous execution occurs.

Content

Scanner excerpt · references/SUPPLEMENTARY_DESIGN.md (reported line 492)May include surrounding context.

md
business_rules:
  confirm_after_review:
    # 审核通过后的自动执行规则
    auto_execute_when:
      risk_level: LOW
      and:
        - review_passed: true

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file's human-readable comments and operator-facing messages are written in Chinese, which can impose a specific language on maintainers or users interacting with this skill configuration. The file does not provide an alternate locale, opt-in mechanism, or justification that this is a region-specific tool.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Automatic cleanup on start, completion, and error is enabled while dry-run is disabled, so deletion behavior can occur without an explicit user-facing checkpoint. This is dangerous because users may trigger routine task execution expecting orchestration only, but instead incur silent destructive actions in the workspace.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The configuration explicitly disables confirmation for deletion actions, including large files, removing an important safety barrier against accidental destructive operations. In combination with broad globbing rules, this increases the likelihood of significant unintended data loss before a user can review the targets.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file uses Chinese-language comments and descriptions throughout, starting with the top-level configuration description, with no indication that language choice is configurable or user-selected. This can violate language/locale policy when a skill or its operational artifacts implicitly force one language without opt-in or justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The automatic pipeline selection is enabled and driven by broad criteria such as complexity_score, task_type, and risk_level without any defined scoring rules, trust boundaries, or validation constraints. In a task-dispatcher skill, this can cause unsafe or incorrect routing of user tasks into pipelines with different approval, review, timeout, and execution behaviors, potentially bypassing stricter handling for risky work or misallocating privileged agents.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This YAML file contains natural-language instructions and labels in Chinese, such as the top-level description and many inline comments, but does not indicate that Chinese is optional or that the skill is intentionally limited to a Chinese-speaking context. Under the language/locale policy, forcing a specific language without user opt-in is a policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

This YAML contains natural-language descriptions primarily in Chinese while operational identifiers and names are partly English, but it does not document any user language preference, opt-in, or locale selection. Under the language/locale policy criterion, this can be treated as an implicit locale choice rather than an explicitly user-selectable one.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

Natural-language comments such as the file description and section labels are presented only in Chinese, which can violate language-choice policy when no opt-in or locale justification is provided. This may make the skill inaccessible or inconsistent for users expecting a different default language.

Content

No source excerpt is available for this finding.