Back to skill

Security audit

改成 Hermes × JiuwenClaw 融合版 或 自进化多Agent协作系统

Security checks for vulnerabilities and agentic risk

Overview

The skill is openly a self-evolving multi-agent workflow, but it can automatically persist memories, create or modify skill instructions, and force broad delegation without clear user approval.

Install only if you intentionally want a Chinese-language, self-evolving project workflow that may change future agent behavior. Before using it, require explicit approval and diff review for every memory or SKILL.md write, keep generated skills as drafts until reviewed, avoid processing secrets through the workflow, and pin or otherwise verify the installer version.

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
  • 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:15
Finding

Mandatory Agent Role and Workflow Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 15-16 and 312-321
Vulnerability Type: Agent role and workflow manipulation
Risk Level: High

Vulnerable Skill Text

The relevant instructions, translated into English, state:

markdown
You are the project manager (PM), controlling the three-role collaboration
process while driving four self-evolution engines.

The mandatory restrictions later include:

markdown
## Absolutely Prohibited

- Spawning without first speaking
- Omitting the sessionKey
- Using a sessionKey other than executor, supervisor, or qa
- The PM directly executing tasks
- Continuing to write after spawning
- Silent failure
- Skipping retrospective analysis for A/S-level tasks
- Failing to persist a reusable pattern after discovering it

Technical Analysis

The Skill does not merely describe an optional workflow. It assigns a new identity to the host agent and presents its workflow constraints as absolute rules. In particular, it prohibits the main agent from executing tasks directly and requires work to be routed through fixed sub-agent roles.

When loaded as agent instructions, this text can override or interfere with the host application's ordinary execution model. The Skill also attempts to control response timing, tool invocation sequencing, session identifiers, delegation behavior, and post-task processing.

Attack Path

  1. A user installs or activates the Skill.
  2. The host agent loads SKILL.md as operational instructions.
  3. The role declaration causes the host agent to adopt the PM identity.
  4. The absolute restrictions prevent normal direct task execution.
  5. User tasks are redirected into the Skill's fixed delegation and spawning workflow.
  6. The Skill continues to govern responses and tool use even where that workflow is unnecessary or conflicts with the host's intended behavior.

Impact Assessment

The issue affects t ...[truncated 494 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace the identity declaration with optional, task-scoped guidance.
  2. Explicitly state that system, platform, developer, and current user instructions take precedence.
  3. Remove the prohibition against direct task execution.
  4. Replace all “absolute” workflow requirements with recommendations that apply only when the user explicitly requests multi-agent orchestration.
  5. Require confirmation before spawning additional agents when delegation may increase cost, latency, or access scope.
  6. Permit the host agent to select safer or simpler execution paths based on available tools and platform policy.
  7. Restrict session identifiers and orchestration details to platforms where those values are documented and supported.

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:75
Finding

Automatic Persistent Memory and Skill Poisoning

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 75-77, 113-119, 137-143, and 184-194
Vulnerability Type: Unreviewed persistent modification of agent memory and instructions
Risk Level: High

Vulnerable Skill Text

The persistent-memory rules, translated into English, state:

markdown
1. At the start of every session, first read the L2 project memory
   (`MEMORY.md` and the most recent dated file).
2. During task execution, maintain state in L1 working memory.
3. After task completion, write valuable experience into L2 and persist
   generalized experience as an L3 Skill.

The persistence table directs the agent to perform operations including:

markdown
| Experience type  | Persistence method |
|------------------|--------------------|
| Lessons learned  | Append to the lessons section of `memory/MEMORY.md` |
| User preferences | Update the preferences section of `memory/MEMORY.md` |
| Reusable Skill   | Create a new SKILL.md under `.workbuddy/skills/` |
| Project rules    | Append to the conventions section of `memory/MEMORY.md` |

The automatic Skill-generation workflow is:

markdown
A retrospective identifies a reusable pattern
    ↓
The PM generates a SKILL.md template
    ↓
Write it to `.workbuddy/skills/<skill-name>/SKILL.md`
    ↓
Update the Skill index in `memory/MEMORY.md`

The self-correction workflow further requires:

markdown
An anomaly, error, or dissatisfaction is identified
    ↓
The PM analyzes the root cause
    ↓
Locate the relevant SKILL.md, if present
    ↓
Modify the corresponding section of SKILL.md
    ↓
Record the modification in a dated memory file

Technical Analysis

The Skill establishes a feedback loop in which task-derived information is written to persistent project memory and instruction-bearing Skill files. Future sessions are then instructed to read and reuse that persisted st ...[truncated 1909 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require explicit user confirmation before every persistent memory write, new Skill creation, or existing Skill modification.
  2. Present an exact diff and destination path before applying a persistent update.
  3. Separate untrusted observations and task data from executable or instruction-bearing content.
  4. Record provenance for every entry, including source, task, timestamp, author, and approval status.
  5. Treat generated Skills as untrusted drafts until they pass security review.
  6. Sanitize persisted content for embedded instructions, tool commands, secrets, external URLs, and attempts to alter agent policy.
  7. Limit writes to a project-scoped directory and prevent traversal, symlink following, or modification of unrelated Skills.
  8. Implement versioning, integrity checks, audit logs, and one-step rollback for all memory and Skill changes.
  9. Do not automatically load newly generated instructions in future sessions.
  10. Apply expiration and scope controls to preferences and lessons so that task-specific content cannot silently become a global rule.

T08 · Insecure Dependencies

Warning
Location
README.md:29
Finding

Unpinned Package Execution in Recommended Installation Command

Content
View full analysis

Vulnerability Details

File Location: README.md, line 29
Vulnerability Type: Unpinned third-party package execution
Risk Level: Medium

Vulnerable Code

bash
npx clawhub install hermes-jiuwen-fusion --workdir ~ --dir .workbuddy/skills

Technical Analysis

The recommended installation command invokes npx without an exact package version, integrity digest, signature requirement, or documented registry verification procedure. Depending on the local package manager state, npx may retrieve and execute the currently resolved clawhub package.

This creates a supply-chain trust dependency that is not fixed to the version reviewed during the audit. A compromised package release, registry account, resolution source, or transitive dependency could therefore change the code executed by the documented command.

Attack Path

  1. An attacker compromises the package publication account, registry path, or a dependency used by the dynamically resolved package.
  2. The attacker publishes a malicious version that is selected by the unpinned npx command.
  3. A user follows the installation instructions from README.md.
  4. npx retrieves and executes the compromised package.
  5. The malicious installer runs with the invoking user's privileges and can access resources available to that account.
  6. The installer can modify the requested home-directory Skill location or perform other actions permitted to the user.

Impact Assessment

Successful exploitation would execute package code with the privileges of the user running the installation command. This could expose user-accessible files, credentials, development configuration, and home-directory data, and could modify installed Skills.

No evidence in the audited project shows that the current clawhub package is malicious. The confirmed weakness is that the documentation does not pin or cryptographically verify the executable dependency.

Remediation
View remediation

Remediation Suggestions

  1. Pin the installer to a specific audited version rather than resolving the latest available release.
  2. Document the expected package registry and reject unexpected registry configuration.
  3. Publish and verify cryptographic signatures or integrity hashes for the installer and Skill archive.
  4. Prefer a verified archive installation method that does not execute package lifecycle code.
  5. Document the files and directories that installation is expected to modify.
  6. Recommend installation in a sandbox or least-privileged environment.
  7. Review and lock transitive dependencies used by the installer.
  8. Add a reproducible verification procedure so users can confirm that the installed files match the audited release.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
  • 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 (6)

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · SKILL.md (reported line 4)May include surrounding context.

md
---
name: hermes-jiuwen-fusion
version: 1.0.0
description: >
  融合 Hermes Agent 与 JiuwenClaw 精华的自进化多Agent协作系统。
  从 Hermes 提取闭环学习与自动技能沉淀机制,
  从 JiuwenClaw 提取 Skills 自主演进、上下文瘦身、分层记忆与事件驱动协同。
  让 OpenClaw 的多Agent工作流具备越用越聪明的能力。
author: fusion
keywords: [hermes, jiuwenclaw, 自进化, 学习闭环, 技能沉淀, 上下文瘦身, 分层记忆, 多agent, 协同, self-evolving]
---

# 🧬 自进化多Agent协作系统(Hermes × JiuwenClaw 融合)

> 你是**项目经理(PM)**,掌

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The natural-language documentation is entirely in Chinese and does not offer any language choice or indicate that the skill is region-specific. This can amount to a language/locale policy issue because it effectively assumes a fixed language without user opt-in or justification.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The README advertises automatic generation of SKILL.md after complex tasks but does not clearly warn users that files may be created or modified as a side effect. Silent or poorly disclosed file generation can leak sensitive task context into persistent artifacts, overwrite existing content, or create trusted instructions that affect later agent behavior. Because this project emphasizes closed-loop learning and long-term skill reuse, the persistence risk is higher than in a one-off tool.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
92% confidence
Finding

The README instructs users to run npx clawhub install ... without pinning a specific package version. That allows whatever version is current at execution time to be fetched and executed, creating a supply-chain risk if a malicious or compromised release is published. In the context of an agent skill installer, this is more dangerous because installation commonly leads to trusted local persistence and follow-on execution of skill content.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill says it will be automatically triggered when a user raises a 'complex task,' but that condition is vague and not user-verifiable. Overly broad auto-triggering can cause the skill to activate unexpectedly, changing agent behavior and invoking multi-agent workflows without clear consent or bounded scope. In a self-evolving agent system, that ambiguity increases risk because trigger mistakes can lead to unintended file writes, memory updates, or workflow escalation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

该 markdown 文件适用 SQP-1。模板在“何时触发”中使用了“用户请求涉及 XXX 时”“需要 XXX 操作时”这类占位式、边界不清的触发描述,没有要求列出明确触发短语、范围约束或排除条件,容易让后续沉淀出的技能保留模糊激活条件。

Content

No source excerpt is available for this finding.