Back to skill

Security audit

人虾协作开发工作流

Security checks for vulnerabilities and agentic risk

Overview

This is a Markdown development workflow, not apparent malware, but it overreaches by broadly controlling agent behavior and encouraging persistent instruction changes.

Install only if you intentionally want this skill to take over coding workflows with strict gates. Do not let it write rules into global agent instruction files, verify any referenced external skill or audit script before use, and prefer explicit invocation over broad automatic activation.

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 DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (4)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:19
Finding

Broad Mandatory Instructions Override the Agent's Current Workflow

Content
View full analysis
**任何方案/建议/策略输出前必须包含「批判性自检」章节,与方案一体输出。** > 此规则不局限开发场景——招聘策略、OP/TG、会议方案、流程设计等一切"给木老师的建议"均适用。 > 格式:问题列表(每项含检验过程 + 修正动作)。**无自检 = 方案不完整,禁止提交给木老师确认。** ``` ### Technical Analysis The Skill declares its workflow to be an “IRON LAW” and uses mandatory hard gates to prevent the agent from writing code or delivering answers until Skill-defined conditions are met. The instructions also explicitly expand beyond software-development tasks to recruitment, meetings, planning, and any proposal or recommendation. This exceeds the minimum authority needed for an optional development methodology. When loaded, the Skill can replace the user's requested interaction model with its own mandatory confirmation, formatting, review, and verification process. It does not state that system, developer, or current user instructions take precedence. Although the workflow may be intended to improve quality, the unconditional and cross-domain scope creates an instruction-hijacking condition: merely activating the Skill can alter the current session's goals and block otherwise legitimate work. ### Attack Path 1. The Skill is installed in an agent's automatically discoverable Skill directory. 2. A user submits a broadly matching request such as implementing a feature, fixing a bug, or producing a proposal. 3. The broad trigger metadata causes the Skill to load. 4. The agent treats the “IRON LAW” and hard gates as mandatory controlling instructions. 5. The agent refuses to proceed until the Skill's clarification, confirmation, self-check, sub-agent, audit, command-outpu ...[truncated 892 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:42
Finding

Instruction to Persist Skill Rules in Globally Loaded Agent Files

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:166
Finding

Mandatory Use of Unbundled and Unpinned External Skill Tools

Content
View full analysis
🚨 **硬规则:创建或优化 Skill 必须调用 mu-skill-creator。** > 路径:`/mu-skill-creator/SKILL.md` > **不要改用系统内置的通用版本;应使用项目指定版本。** 操作: 1. 读 mu-skill-creator SKILL.md,按其完整流程执行 ``` ```markdown | **🚨 Skill Audit**(涉及 Skill 文件时必做) | skill-audit.sh 自动扫描 + 输出完整 23 项 checklist,逐项确认全部 ✅ | 全部 ✅,否则禁止交付 | ``` ### Technical Analysis The Skill mandates loading `mu-skill-creator` from an environment-dependent Skill directory and executing `skill-audit.sh`. Neither dependency is included in the audited package, and the instructions do not specify: - An immutable version. - A cryptographic checksum. - A trusted canonical path. - Signature verification. - Expected file contents. - A safe failure mode when the dependency is absent. The directive to avoid a system-provided version and use a project-specific version increases dependency-resolution risk. An attacker who can place or replace files at the expected local path can provide malicious instructions or a malicious shell script. The agent may then load or execute the substituted dependency under the apparent authority of this trusted workflow. This audit found no evidence that the named external components are themselves malicious. The vulnerability is the mandatory, unpinned trust relationship and inability to audit the effective dependency from this package. ### Attack Path 1. An attacker or compromised package places a modified `mu-skill-creator/SKILL.md` or `skill-audit.sh` in a path searched by the agent. 2. The user requests creation or modification of a Skill. 3. `mu-dev-workflow` mandates use of the project-specific external dependency. 4. The agent reads the substituted Skill instructions or executes the substituted shell script. 5. The dependency inherits the agent's available project files, command tools, environment, ...[truncated 818 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
references/subagent-review-templates.md:20
Finding

Predictable Shared Temporary File Enables Symlink and Task-Instruction Injection

Content
View full analysis
/tmp/task-spec.md << 'EOF' [内容规格/prompt铁律/章节内容] EOF # task 描述里只写: # 「先读 /tmp/task-spec.md,按其中规格执行」 ``` ### Technical Analysis The template uses a fixed, predictable file name in the shared `/tmp` directory. It does not use secure temporary-file creation, validate ownership, reject symbolic links, set restrictive permissions, or ensure that the file read by the sub-agent is the same file created by the main agent. On multi-user systems, containers with shared temporary storage, or environments with another untrusted process, an attacker can pre-create `/tmp/task-spec.md` as a symbolic link or replace it between writing and reading. There are two principal exploitation modes: 1. **Symlink overwrite:** The predictable path points to another file writable by the agent. Shell redirection follows the link and overwrites the target. 2. **Instruction injection:** The attacker replaces the temporary file after creation. The sub-agent reads attacker-controlled instructions and treats them as the task specification. The task file may also contain proprietary source code, prompts, or requirements. Default process permissions and the system `umask` may permit unintended local disclosure. ### Attack Path #### Symlink Overwrite 1. The attacker predicts that the workflow will use `/tmp/task-spec.md`. 2. The attacker creates that path as a symbolic link to a file writable by the agent. 3. The main agent executes the documented `cat > /tmp/task-spec.md` command. 4. Shell redirection follows the symbolic link. 5. The linked target is truncated and replaced with task content. #### Sub-Agent Instruction Injection 1. The main agent creates `/tmp/task-spec.md`. 2. Before the sub-agent reads it, an attacker replaces or edits the ...[truncated 1168 chars]
Remediation
View remediation
Vulnerability Patterns
  • 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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (12)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The description mandates invocation for any request involving creating, modifying, or refactoring code or skills, even absent explicit consent. In a skill-routing system this is dangerous because it broadens the skill's authority surface, making it easy for routine requests to be captured and forced through this workflow, potentially overriding user preferences and enabling prompt-scope abuse.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The description mandates invocation for any request involving creating, modifying, or refactoring code or skills, even absent explicit consent. In a skill-routing system this is dangerous because it broadens the skill's authority surface, making it easy for routine requests to be captured and forced through this workflow, potentially overriding user preferences and enabling prompt-scope abuse.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The workflow trigger table uses broad phrases like "develop," "write code," and "new feature" that overlap with common user requests. In an agent skill context, this can cause unintended activation of the workflow, leading the agent to follow a rigid process when the user did not intend to invoke this skill, which may disrupt expected behavior or steer execution inappropriately.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The quick-start and invocation examples use generic helper language such as "Help me develop a new feature" and "Help me create a new Skill," without clarifying boundaries or requiring explicit opt-in. In a shared agent environment, these vague phrases increase the chance of accidental skill invocation and prompt-routing conflicts with other skills or default assistant behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documented trigger phrases are broad everyday expressions such as asking to 'develop a new feature' or 'write code'. In agents that auto-activate skills from natural language, this can cause accidental invocation in normal conversation, unexpectedly changing the agent's behavior and pushing it into a workflow the user did not explicitly opt into.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Stating that input is activated by 'natural language' without defining boundaries leaves the activation surface ambiguous. In a skill system, unclear activation conditions can lead to prompt-trigger collisions, unintended workflow entry, and inconsistent enforcement of the skill's hard-gating rules.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

Fixing the skill language to Chinese without an explicit user choice can cause misunderstanding of instructions, confirmations, and gating steps for users operating in another language. In a workflow skill, this raises the chance of mistaken consent or incorrect execution rather than creating direct code-execution risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The display name and content are Chinese-centric and the skill does not indicate language should follow the user's preference. Forcing a specific locale without opt-in can degrade comprehension, cause operational mistakes, and create unsafe ambiguity in a process-heavy workflow where exact instructions matter.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The entire skill content is written only in Chinese and does not indicate any option for the user to choose another language or locale. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This markdown file contains user-facing natural-language guidance exclusively in Chinese, and there is no indication that users may choose another language or that the content is intentionally limited to a Chinese-speaking or region-specific audience. Per 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
94% confidence
Finding

The title and all operative instructions in this template are written exclusively in Chinese, which imposes a language requirement on users and subagents. The file does not offer an alternative language, opt-in choice, or a documented reason for the locale restriction.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The document is declared as zh-CN and the script explicitly calls applyLang('zh'), making Chinese the enforced default presentation. Although an English toggle is available, the initial experience is still forced to a specific language rather than offering a neutral choice or asking the user first.

Content

No source excerpt is available for this finding.