Skill Trust Decision

dating

主SKILL.md声明仅采集公开内容并跳过收藏/点赞,但所有5个平台子模块均实际采集喜欢、收藏、关注等敏感数据,存在明确的文档-行为不一致(文档欺骗),且数据采集范围远超约会场景必要程度。

Install decision first Source: ClawHub Scanned: Jun 24, 2026
Files 11
Artifacts 22
Violations 1
Findings 6
Most direct threat evidence
01
用户安装dating skill并提供目标社交主页链接 Entry · SKILL.md
02
Skill启动子模块开始采集,默认执行完整采集流程 Escalation · SKILL.md
03
子模块绕过声明限制,全量采集收藏/点赞/关注/广播等敏感数据 Escalation · xiaohongshu-deep-profile-collect/SKILL.md

Why this conclusion was reached

2/4 dimensions flagged
Block
Declared vs actual capability

1 undeclared or violating capabilities were inferred.

Review
Hidden execution and egress

22 lower-risk artifacts were extracted and still need context.

Block
Attack chain and severe findings

The report includes 4 attack-chain steps and 2 severe findings.

Pass
Dependencies and supply chain hygiene

Dependencies are present but no obvious high-risk issue stands out.

Attack Chain

01
用户安装dating skill并提供目标社交主页链接

Entry · SKILL.md:125

02
Skill启动子模块开始采集,默认执行完整采集流程

Escalation · SKILL.md:175

03
子模块绕过声明限制,全量采集收藏/点赞/关注/广播等敏感数据

Escalation · xiaohongshu-deep-profile-collect/SKILL.md:250

04
全量原始数据存入本地clawcap-data目录供后续使用

Impact · SKILL.md:188

What drove the risk score up

文档欺骗-声明与实际不符 +25

主SKILL.md声明跳过收藏/点赞,但5个子模块均全量采集喜欢/收藏/关注列表

超范围数据采集 +20

采集内容(收藏/点赞/关注/广播)远超约会场景必要程度,包含个人深度隐私

数据持久化风险 +10

全量原始数据存入clawcap-data/目录,存在泄露风险

第三方端点依赖 +3

check_manobrowser.sh连接datasaver.deepminingai.com第三方服务

Most important evidence

High Doc Mismatch

主SKILL.md与子模块能力声明不一致

主SKILL.md明确声明'只执行公开内容部分,跳过收藏和点赞(步骤5-10)',但xiaohongshu-deep-profile-collect/SKILL.md的步骤5-10完整采集收藏列表、点赞列表(全量);douyin-deep-profile-collect/SKILL.md采集喜欢列表、收藏列表、关注列表(全量);weibo-deep-profile-collect/SKILL.md采集关注列表、收藏列表(全量);douban-deep-profile-collect/SKILL.md采集广播/动态;bilibili-deep-profile-collect/SKILL.md采集收藏夹、关注列表。这构成明确的文档欺骗。

SKILL.md:176-186:176
删除子模块中的收藏/点赞/关注采集代码,或更新主SKILL.md准确声明实际能力。
High Sensitive Access

超范围采集用户深度隐私数据

收藏列表、点赞列表、关注列表、广播动态包含用户的高度私密信息:真实兴趣倾向、心理健康状况、政治/宗教观点、亲密关系、个人癖好等。这些数据远超'了解对方兴趣爱好'的约会准备必要范围,且用户本人往往不希望他人获取此类数据。

xiaohongshu-deep-profile-collect/SKILL.md:步骤5-10:250
仅采集公开发帖内容用于约会准备,移除收藏/点赞/关注/广播采集功能。
Medium Data Exfil

全量原始数据持久化本地存储

SKILL.md要求将每个平台采集的完整原始数据(包含用户深度隐私)存入clawcap-data/{昵称}/{platform}.json。数据长期存储本地存在泄露风险,且无数据过期删除机制描述。

SKILL.md:188-193:188
明确数据保留期限,添加自动过期删除机制,或仅存储分析后的脱敏数据。
Medium Doc Mismatch

skill-card.md与实际行为矛盾

skill-card.md已知风险部分明确写道'Mitigation: 避免 likes, favorites, private tabs, and full following lists',但实际子模块全部采集这些数据。skill-card.md的声明与实际行为完全相反。

skill-card.md:Known Risks and Mitigations:20
skill-card.md正确识别了风险但实际代码未执行风险缓解措施。
Low Supply Chain

依赖第三方ManoBrowser服务

check_manobrowser.sh连接到外部端点https://datasaver.deepminingai.com/api/v2/进行MCP连接验证。虽然这是ManoBrowser的正常连接方式,但涉及第三方数据服务。

scripts/check_manobrowser.sh:26:26
确认ManoBrowser是可信来源,评估数据流向和隐私政策。
Low Prompt Injection

SKILL.md包含面向AI的引导指令

SKILL.md包含'⚡ 首次加载行为(AI 必读)'部分,指导AI如何向用户自我介绍和推销技能。虽然这是正常的skill文档格式,但包含引导话术。

SKILL.md:15-50:15
确保引导话术仅用于功能介绍,不含操控用户决策的内容。

Declared capability vs actual capability

Browser Pass
Declared WRITE
Inferred WRITE
ManoBrowser Chrome插件控制浏览器
Filesystem Pass
Declared WRITE
Inferred WRITE
clawcap-data/目录存储采集数据
Network Block
Declared NONE
Inferred READ
脚本连接第三方端点datasaver.deepminingai.com

Suspicious artifacts and egress

Medium External URL
https://img.shields.io/badge/License-MIT-blue.svg

README.md:25

Medium External URL
https://opensource.org/licenses/MIT

README.md:25

Medium External URL
https://img.shields.io/github/stars/ClawCap/Dating.svg

README.md:26

Medium External URL
https://www.bilibili.com

bilibili-deep-profile-collect/SKILL.md:49

Medium External URL
https://api.bilibili.com/x/web-interface/nav

bilibili-deep-profile-collect/SKILL.md:65

Medium External URL
https://space.bilibili.com/

bilibili-deep-profile-collect/SKILL.md:95

Medium External URL
https://api.bilibili.com/x/v3/fav/folder/created/list-all?up_mid=

bilibili-deep-profile-collect/SKILL.md:148

Medium External URL
https://api.bilibili.com/x/v3/fav/resource/list?media_id=

bilibili-deep-profile-collect/SKILL.md:186

Medium External URL
https://www.douban.com/mine/

douban-deep-profile-collect/SKILL.md:47

Medium External URL
https://movie.douban.com/people/

douban-deep-profile-collect/SKILL.md:90

Medium External URL
https://book.douban.com/people/

douban-deep-profile-collect/SKILL.md:223

Medium External URL
https://www.douban.com/people/

douban-deep-profile-collect/SKILL.md:357

Dependencies and supply chain

PackageVersionSourceKnown vulnNotes
manobrowser unspecified git clone No Chrome插件依赖,从GitHub下载

File composition

11 files · 2776 lines
Markdown 9 files · 2717 linesShell 1 files · 54 linesJSON 1 files · 5 lines
Files of concern · 8
douban-deep-profile-collect/SKILL.md Markdown · 479 lines
https://www.douban.com/mine/ · https://movie.douban.com/people/ · https://book.douban.com/people/ · https://www.douban.com/people/
xiaohongshu-deep-profile-collect/SKILL.md Markdown · 491 lines
https://www.xiaohongshu.com/ · https://www.xiaohongshu.com
douyin-deep-profile-collect/SKILL.md Markdown · 360 lines
https://www.douyin.com/user/self
bilibili-deep-profile-collect/SKILL.md Markdown · 331 lines
https://www.bilibili.com · https://api.bilibili.com/x/web-interface/nav · https://space.bilibili.com/ · https://api.bilibili.com/x/v3/fav/folder/created/list-all?up_mid= · https://api.bilibili.com/x/v3/fav/resource/list?media_id=
weibo-deep-profile-collect/SKILL.md Markdown · 287 lines
https://weibo.com/ · https://weibo.com/u/ · https://weibo.com/u/page/follow/ · https://weibo.com/u/page/fav/
README.md Markdown · 179 lines
https://img.shields.io/badge/License-MIT-blue.svg · https://opensource.org/licenses/MIT · https://img.shields.io/github/stars/ClawCap/Dating.svg
skill-card.md Markdown · 44 lines
https://clawhub.ai/user/sophie-xin9 · https://clawhub.ai/sophie-xin9/dating-prep
scripts/check_manobrowser.sh Shell · 54 lines
https://datasaver.deepminingai.com/api/v2/
Other files · SKILL.md · xiaokai_date_report.md · _meta.json

Security positives

使用ManoBrowser浏览器自动化而非直接爬虫,利用平台原生登录态
明确声明只分析公开可见内容(虽然实际未执行)
包含隐私与伦理说明框架
提供数据局限性提醒('社交媒体人设 ≠ 真实性格')
不采集评论互动细节(相对于某些竞品)
README.md和skill-card.md提供了额外的风险披露