Back to skill

Security audit

CGArtLab Quill

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious with medium confidence, mainly because it reports guidance to inspect broad environment-variable namespaces that may expose credentials; Skillspector coverage was partial.

Consider the credential exposure risk before installing. Skillspector recommends not installing, and its scan was partial: it covered 36 of 37 components, with one file only partly inspected and tool-misuse analysis degraded. The other scanner completed with no findings, but that does not resolve the reported concern or make the evidence complete.

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
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (8)

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · LESSONS.md (reported line 63)May include surrounding context.

md
## 发布与安全扫描

- 失误:ClawHub 直接发整个仓库,`evals/run.mjs` 的 `spawnSync` 被判 critical、研究笔记里当反面教材引用的 `curl | bash` 被算成行为证据 → 规则:registry 发布包与仓库分离,先打包再发(排除 `evals/` `docs/` dev-doc),排除清单只写在 `.github/actions/publish` 打包步骤那一处
- 失误:`unchanged` 断言只遍历跑后仍存在的文件,草稿被删掉也判「逐字节未变」 → 规则:关于「还在不在」的断言必须双向遍历(快照→现在查删除与改动,现在→快照查新增),并配自测 `evals/selftest-unchanged.mjs`
- 失误:`evals/run.mjs` 在 agent 跑完之后才拍快照 → 规则:快照必须在被测动作之前拍,晚一步就等于拿结果跟结果比
- 失误:披露模板默认带「会话链接」 → 规则:会把对话内容交给第三方的模板默认不写,需要时由作者确认接收方能看到的全部内容

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · CHANGELOG.md (reported line 87)May include surrounding context.

md
<summary>技术细节:为什么确定是这一步造成的</summary>

- `git diff v0.11.0..HEAD`:除第零步(读作者)外,`references/writing-craft.md`、`references/style-and-bans.md`、`references/form-guide.md` **一行未动**;其余改动都是安全/发布/文档层面的。
- 本机 Magpie 安装目录 `~/.config/magpie/library/skills/quill/SKILL.md` 实测就是 0.11.2(与仓库 diff 一致)——用户最近几次用的正是这版。
- 直接冲突点:第零步改成「没同意不读任何历史文章」,第七步仍要求「读其已发表文章」取声音指纹。两条同时在场时,声音校准必然落空。

</details>

Environment detection instructs the agent to read broad environment-variable namespaces without warning that they may contain credentials.

Medium
Category
Output Handling
Confidence
89% confidence
Finding

Reading entire environment-variable namespaces can bring API keys, tokens, or sensitive configuration into the model context and tool logs unnecessarily. Harness detection needs only limited presence checks, and the document's safeguards for credential-bearing configuration files do not extend to these variables; no external exfiltration is demonstrated.

Content

No source excerpt is available for this finding.

The title-writing procedure requires Chinese output without user language opt-in.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The instruction explicitly requires writing the complete intended meaning in Chinese before shortening it. This imposes a particular natural language, rather than merely requesting clear wording, and is not conditional on the user's chosen article language.

Content

No source excerpt is available for this finding.

The form guide explicitly extends Quill to short answers and social content despite the manifest's long-form-only scope.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

The manifest limits activation to argumentative nonfiction long-form writing and excludes short texts such as comments and instant messages. However, L168 directs Quill to produce short answers using its writing rules without the project workflow, and L172–L176 provide a separate short-form/social-content mode rather than restricting assistance to long-form work.

Content

No source excerpt is available for this finding.

The image-compression step recommends TinyPNG without warning that using it uploads images to a third-party service.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow instructs the agent to compress oversized images using TinyPNG or other tools, but does not distinguish remote processing from local compression. Images may contain private information, and the subsequent instruction to obscure sensitive details does not warn about disclosure during compression or require redaction before upload.

Content

No source excerpt is available for this finding.

L237 requires Chinese-locale typography by default without checking the user's preferred language or locale.

Low
Category
Output Handling
Confidence
90% confidence
Finding

The validation checklist requires Chinese typography, including full-width punctuation and corner quotation marks. Unlike the preceding platform-specific requirement, this default is not restricted to Chinese-language articles or a justified regional publishing context, so it can impose inappropriate formatting on other-language work.

Content

No source excerpt is available for this finding.

The writing workflow provides handling for poetry and scripts despite the manifest explicitly excluding those forms.

Low
Category
Excessive Agency
Confidence
93% confidence
Finding

At L548, the workflow instructs the agent to handle poetry, scripts, dialogue, and experimental writing using relaxed checks rather than treating excluded forms as outside the skill. This expands the operational scope beyond the manifest’s explicit exclusion of poetry and scripts from its nonfiction long-form writing purpose.

Content

No source excerpt is available for this finding.