Back to skill

Security audit

dsh-capability-loop

Security checks for vulnerabilities and agentic risk

Overview

The skill is marked suspicious because scanners report potentially destructive cleanup, unsafe package contents, and advice to weaken antivirus protection; some files were only partially analyzed.

Treat this as a review-needed skill and consider avoiding installation until the directory-deletion behavior and package contents are checked, and the antivirus guidance is reviewed. Coverage was partial: 37 of 40 components were fully inspected, with three partially inspected and none entirely uninspected, so the scan does not establish that all behavior is safe.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (20)

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 10)May include surrounding context.

md
> **版本**:`version.json`(权威)+ `CHANGELOG.md`(每次写回必须 bump)+ `SKILL.md` frontmatter 的 `version`(必须与 `version.json` 一致;技能市场用它显示「已安装版本」)。**首次使用先读 `QUICKSTART.md`**(1

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 45)May include surrounding context.

md
3. **写回本 skill 必须版本化**:改 `references/`/`scripts/`/`SKILL.md` 任意一处,同一提交内更新 `version.json` + `CHANGELOG.md`,并至少跑 1 正例 + 1 反例回归。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 142)May include surrounding context.

md
- 每次 bump 在 `CHANGELOG.md` 写:日期、版本、改了什么、为什么、验证方式、回滚方式。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 160)May include surrounding context.

md
- 版本对照:对方 `version.json` 与你不同不是问题;用 `CHANGELOG.md` 做差异说明,不做自动合并。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 10)May include surrounding context.

md
> **版本**:`version.json`(权威)+ `CHANGELOG.md`(每次写回必须 bump)+ `SKILL.md` frontmatter 的 `version`(必须与 `version.json` 一致;技能市场用它显示「已安装版本」)。**首次使用先读 `QUICKSTART.md`**(1

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 45)May include surrounding context.

md
3. **写回本 skill 必须版本化**:改 `references/`/`scripts/`/`SKILL.md` 任意一处,同一提交内更新 `version.json` + `CHANGELOG.md`,并至少跑 1 正例 + 1 反例回归。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 152)May include surrounding context.

md
**拆分规则**:`SKILL.md` 正文 **< 300 行**;任一 reference **≤ 200 行**;超了就拆文件并在 §7 地图登记;拆完后 `name` 不变。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 126)May include surrounding context.

md
| 新坑/新错误/新规避 | `references/70-playbook.md` | `ERR-YYYYMMDD-NN`:现象 / 根因 / 修法 / 是否已复现 |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 177)May include surrounding context.

md
| `references/70-playbook.md` | 失败手册:PS5.1/BOM/反引号/crashpad/索引/端口/编码/多帧 zstd/MCP 等 |

The smoke test promises temporary-only changes but recursively deletes any supplied Scratch directory.

High
Category
Excessive Agency
Confidence
100% confidence
Finding

Scratch is caller-controlled and is passed directly to recursive forced deletion without checking ownership, containment, or whether it contains existing project data. Despite the temporary-only promise, supplying a real project or workspace path can destroy its contents; the final cleanup repeats the same unsafe operation.

Content

No source excerpt is available for this finding.

L067 recursively deletes an existing Scratch directory without disclosing that its pre-existing contents will be lost.

High
Category
Output Handling
Confidence
100% confidence
Finding

The initialization deletes all pre-existing contents of the supplied Scratch directory without warning or confirmation. This is the same underlying destructive-path defect as the other finding, and even -Keep does not prevent the initial deletion.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Recommends antivirus exclusions without warning that they reduce malware protection.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The guide recommends antivirus exclusions for node.exe and the application executable without explaining the protection lost or establishing antivirus interference as the cause. Excluding a general-purpose script runtime can reduce detection of malicious scripts and child-process activity, depending on the security product's exclusion semantics.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/70-playbook.md (reported line 73)May include surrounding context.

md
**ERR-20261112-11 · 沙箱里 Chrome/agent-browser 起不来**
现象:`Chrome exited early (exit code: -36863) without writing DevToolsActivePort`;stderr:`crashpad_client_win.cc:431 OpenProcess: 拒绝访问 (0x5)`。
根因:DSH workspace-write 进程沙箱拒绝 crashpad;`~/.agent-browser` 也不可写。
修法:不要反复重试;走 A/B/C 路径(全权限会话 / 用户终端截图 + `read_image` / 手动截图标注未实测);若要重定向状态目录,`AGENT_BROWSER_SOCKET_DIR/PROFILE/SCREENSHOT_DIR` 指到工作区(但 crashpad 问题仍在)。

**ERR-20261112-12 · 截图验证靠「肉眼看」没有留证**

The plugin-removal procedure does not warn about losing persisted plugin data.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

L083 instructs the agent to delete configuration entries, uninstall the plugin, and remove its storage. It does not explain that removing storage may discard saved settings or other persisted plugin data, or require a backup before proceeding.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/95-store-publish.md (reported line 24)May include surrounding context.

md
| 市场上架 skill 的安装目录为 `<handle>--<slug>` | 本机 `~/.dsh/skills/clawhub-pskoett--self-improving-agent/` 等 |
| ClawHub **安装版**目录含 `_meta.json`(发布是否同写未实测) | 本机 `clawhub-halthelobster--proactive-agent/_meta.json` = `{"ownerId","slug","version","publishedAt"}` |
| 市场「已安装版本」取自 `SKILL.md` frontmatter `version` | 插件 `lib/index.js:1493-1504` `declaredVersionOf`:frontmatter 无 `version` 时为空(v2.2.0 本地导入实测 `version=` 空);v2.2.1 起 frontmatter 声明 `version`,本地导入实测 `version=2.2.1` |
| SkillHub 的 clawhub 来源与自动同步 | `https://api.skillhub.cn/api/skills`:`source:"clawhub"`、`upstream_url:"https://clawhub.ai/<user>/<slug>"`、`namespace.canonicalName:"@clawhub_<user>/<slug>"`;详情接口 `latestVersion.changelog:"Synced by skillhub pipeline"` |
| 发布命令/登录方式 | `clawhub` npm 包 README(`https://registry.npmmirror.com/clawhub`):`clawhub login`、`clawhub skill publish <dir> --slug ... --name ... --changelog ...`;Windows 配置在 `%APPDATA%\clawhub\config.json` |

## 3. 发布前准备(本 skill 已内置校验)

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/95-store-publish.md (reported line 75)May include surrounding context.

md
1. SkillHub 由同步管线从 ClawHub 拉取(证据见 §2),**没有由你手动触发的入口**;同步时延未实测(§6)。
2. 同步后在 `https://skillhub.cn/skills` 搜索 slug,或直接查 API:
   `https://api.skillhub.cn/api/v1/skills/<slug>`(返回含 `source:"clawhub"`、`upstream_url`)。
3. 打开 DSH **Skill Market** 面板 → 搜索 → 安装;安装到 `~/.dsh/skills/<handle>--dsh-capability-loop/`。
4. **重启 DeepSeek Harness**,新会话即可调用。
5. 之后 `skill-status.ps1` 能读到面板里的 `latestVersion`,版本落后时只提示「去市场更新」。

Generated publishing notes claim sensitive material is excluded, but packaging copies every enumerated source file.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The script copies every enumerated source file while its generated notes assert that credentials, sessions, storage, and backups are excluded. Sensitive-content checks cover only selected extensions and merely warn, so private keys, session files, or backups can enter a distributable package and subsequently be published under a false assurance of exclusion.

Content

No source excerpt is available for this finding.

The visual QA probe is documented as read-only but creates persistent files and modifies a browser profile directory.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

L001 explicitly describes this script as read-only detection, but L019 creates an output directory and L062–L074 write a test HTML page and launch Chrome to create a screenshot and browser profile. L112–L113 then recursively delete the temporary profile while retaining the HTML/PNG, so the probe has filesystem side effects contrary to its documented read-only guarantee.

Content

No source excerpt is available for this finding.

Instructs the agent to persist inferred user preferences without a user-facing privacy notice.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The template directs the agent to infer preferences from user feedback and persist them in memory/user-profile.md without specifying notice, consent, or limits on sensitive information. This can retain private or inaccurate inferences beyond the immediate task and influence future interactions, especially if workspace memory is shared or version-controlled. Preference learning fits the skill’s improvement workflow, making benign intent plausible, but the template provides no privacy safeguards; no external transmission is shown.

Content

No source excerpt is available for this finding.