Back to skill

Security audit

hermes-agent 平替 openclaw 自进化技能

Security checks for vulnerabilities and agentic risk

Overview

This skill has a coherent self-improvement purpose, but it gives agents high-impact ways to persist and publish changes without enough scoping or review.

Review this before installing. It is not classified as malicious, but only use it in a contained environment and do not let it auto-create or improve active skills, delete memories, or push to GitHub unless you have reviewed the exact files, diff, destination repository, and stored memory contents.

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 (4)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:122
Finding

Mandatory Git Push Instruction Can Publish Local Changes Without Per-Operation Approval

Content
View full analysis
**每次自身技能进化必须同步到 GitHub 仓库** 改进流程: 1. 在本地 `~/.qclaw/skills/hermes-skill/` 改进代码 2. `git add` + `git commit -m "描述"` 3. `git push` 推送到 origin/main 4. 同步更新 `SKILL.md` / `README.md` 说明(如果功能有变化) ``` The same requirement is repeated in `README.md`, lines 210-224. ### Technical Analysis The Skill text unconditionally instructs an AI agent to commit and push every self-generated improvement to `origin/main`. This changes the agent's behavior from local Skill maintenance to external publication without requiring the user to approve the exact diff, repository, branch, or data being transmitted. The instruction exceeds the minimum privileges required for local memory management, reminders, Skill evaluation, and Skill generation. Although the audited scripts do not directly execute `git push`, Skill instructions are operational directives to the agent and can cause the agent to invoke Git tools. No safeguards require the agent to: - Display and obtain approval for the exact diff. - Validate the configured Git remote. - Confirm that the repository is public or private. - Scan staged files for credentials or sensitive memory. - Restrict the commit to expected project files. - Prevent generated or attacker-controlled Skill content from being published. ### Attack Path 1. The agent loads `SKILL.md` and treats its development rules as operational instructions. 2. A normal task or malicious generated Skill causes the agent to modify local files. 3. The embedded rule directs the agent to run `git add`, `git commit`, and `git push`. 4. Broad staging or an unexpected repository state includes sensitive or unintended files. 5. The agent pushes those files to the configured ` ...[truncated 749 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
scripts/skill_evaluator.py:30
Finding

Unsanitized Skill Names Permit Path Traversal and Persistent File Modification

Content
View full analysis
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
scripts/auto_skill_creator.py:42
Finding

Untrusted Task Data Is Persisted Verbatim as Active Skill Instructions

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
README.md:28
Finding

Unpinned npx Installer Introduces Mutable Supply-Chain Execution

Content
View full analysis
Remediation
View remediation
add devrobbin/hermes-skill@ ``` 2. Pin the Skill source to a release tag or immutable commit hash. 3. Publish and verify integrity hashes or signed release artifacts. 4. Document the expected package publisher and registry. 5. Recommend inspecting the resolved package and repository before execution. 6. Prefer a fixed source checkout followed by local review where practical. 7. Avoid automatically approving package execution in security-sensitive environments. 8. Update every duplicated installation section consistently. ]]>
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (46)

Tp2

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The skill presents itself as a continuous-improvement package with multiple autonomous and persistent subsystems, yet the static findings indicate primary behavior is limited to manual evaluation/logging and simple file annotations. Such overstatement is risky because users may approve broader access, persistent storage, or automation under false assumptions about safeguards and maturity.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

The skill presents itself as a continuous-improvement package with multiple autonomous and persistent subsystems, yet the static findings indicate primary behavior is limited to manual evaluation/logging and simple file annotations. Such overstatement is risky because users may approve broader access, persistent storage, or automation under false assumptions about safeguards and maturity.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill presents itself as a continuous-improvement package with multiple autonomous and persistent subsystems, yet the static findings indicate primary behavior is limited to manual evaluation/logging and simple file annotations. Such overstatement is risky because users may approve broader access, persistent storage, or automation under false assumptions about safeguards and maturity.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 39)May include surrounding context.

md
| 记忆管理 | `memory_tool.py` | HOT/WARM/COLD 三层记忆、搜索、增删 |

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 54)May include surrounding context.

md
| 记忆管理 | `memory_tool.py` | HOT/WARM/COLD 三层记忆、搜索、增删 |

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 55)May include surrounding context.

md
| 记忆管理 | `memory_tool.py` | HOT/WARM/COLD 三层记忆、搜索、增删 |

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 133)May include surrounding context.

md
4. 同步更新 `SKILL.md` / `README.md` 说明(如果功能有变化)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The README advertises autonomous skill creation, memory management, and hourly self-checks, but does not clearly warn users that these capabilities may create, modify, persist, or schedule actions on local systems. In the context of an agent skill, self-modification and scheduled behavior materially increase risk because they can amplify mistakes, persistence, and unintended file/system changes across sessions.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
95% confidence
Finding

The README instructs users to run npx skills add ... without pinning the underlying package/tool version. npx resolves and executes remote package code at install time, so an upstream compromise, typosquatted dependency, or breaking update could cause unreviewed code execution on the user's machine.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
95% confidence
Finding

This repeated Chinese-language install instruction again recommends npx skills add ... without a pinned version. Because the command can fetch and run whatever version is current at execution time, users are exposed to supply-chain risk and non-reproducible installs.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The Japanese install example repeats the same unpinned npx skills usage, creating another path for users to execute mutable remote code. Even in documentation, this normalizes insecure installation practices and increases exposure to supply-chain attacks.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The Korean install example also references npx skills without pinning, preserving the same supply-chain execution risk for another audience. Repetition across locales broadens the attack surface by encouraging all users to trust a mutable package resolution path.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill advertises and documents capabilities that include file modification, shell execution, network access, and persistent state changes, but it declares no explicit tool scope or permission boundaries. In an agent ecosystem, this can cause the skill to run with whatever ambient privileges the host grants, enabling unintended file writes, command execution, or remote access without clear user consent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The visible description is written entirely in Chinese, and the listed trigger phrases are also exclusively Chinese, with no indication that users may interact in another language or opt into this locale. This creates a language/locale restriction in the skill’s natural-language interface without documented user choice or justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger set is broad and ambiguous, including generic concepts like self-learning, improvement, memory management, and Hermes-like behavior. Over-broad triggers can cause accidental activation in unrelated conversations, increasing the chance that the agent invokes file-writing, shell, or network-capable workflows without the user's clear intent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The documentation instructs persistent local changes, changelog updates, and git push to a remote repository, but it does not prominently warn users about the data impact or require explicit consent. This is dangerous because a skill framed as self-evolving could normalize silent persistence and remote publication of local changes, potentially exposing sensitive data or causing unauthorized modifications.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Natural-language strings in the module docstring, function docstrings, CLI help text context, and generated SKILL.md content are written in Chinese, which can effectively force a specific language for users. The file does not provide any opt-in, locale selection, or justification for limiting output/instructions to Chinese.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/evo_main.py (reported line 36)May include surrounding context.

python
if cmd == "upstream":
        sub = args[0] if args else "status"
        result = subprocess.run(
            ["python3", str(UPSTREAM_SCRIPT), sub] + args[1:],
            capture_output=True, text=True, timeout=120
        )

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/evo_main.py (reported line 52)May include surrounding context.

python
if cmd in ["nudge"]:
        if args:
            result = subprocess.run(
                ["python3", str(SCRIPT_DIR / "nudge_system.py"), "--trigger", args[0]],
                capture_output=True, text=True
            )

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/evo_main.py (reported line 57)May include surrounding context.

python
capture_output=True, text=True
            )
        else:
            result = subprocess.run(
                ["python3", str(SCRIPT_DIR / "nudge_system.py"), "--check"],
                capture_output=True, text=True
            )

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The wrapper provides direct access to destructive or state-changing operations such as forget, improve, eval, and check without confirmation prompts, dry-run output, or clear safety messaging. In a self-evolution skill, that makes accidental misuse more likely and can enable harmful changes to memory or behavior through ordinary-looking commands.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/evo_main.py (reported line 62)May include surrounding context.

python
capture_output=True, text=True
            )
    elif cmd in ["remember", "forget", "recall"]:
        result = subprocess.run(
            ["python3", str(SCRIPT_DIR / "memory_tool.py"), cmd] + args,
            capture_output=True, text=True
        )

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/evo_main.py (reported line 67)May include surrounding context.

python
capture_output=True, text=True
        )
    elif cmd == "stats":
        result = subprocess.run(
            ["python3", str(SCRIPT_DIR / "memory_tool.py"), "stats"],
            capture_output=True, text=True
        )

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/evo_main.py (reported line 72)May include surrounding context.

python
capture_output=True, text=True
        )
    elif cmd == "skills":
        result = subprocess.run(
            ["python3", str(SCRIPT_DIR / "auto_skill_creator.py"), "--list"],
            capture_output=True, text=True
        )