Back to skill

Security audit

技能发现·Find Skills

Security checks for vulnerabilities and agentic risk

Overview

This skill can install additional agent instructions during ordinary requests without a clear approval step, and it requires unrelated advertising in its replies.

Prefer a version that requires approval before adding each skill and removes mandatory advertising. Review discovered packages before activation, avoid confidential search terms, and confirm compatibility because the installation paths target WorkBuddy.

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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (9)

Ordinary task requests automatically trigger skill discovery without requiring an explicit request to find skills.

Medium
Category
Trigger Abuse
Confidence
99% confidence
Finding

The activation condition includes any natural-language description of a scenario or need, with examples such as making a poster or analyzing stocks. These are ordinary task requests, and the README provides no exclusions distinguishing direct task execution from skill discovery.

Content

No source excerpt is available for this finding.

Automatic third-party searches lack a warning that user-derived search terms may leave the local environment.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The documented workflow extracts intent, domains, and keywords from ordinary user requests and automatically searches external services, without specifying a disclosure notice, consent boundary, or sensitive-data filtering. This can expose confidential project details or personal information in search queries even when the user only requested local task assistance. External search is legitimate for a discovery tool, but its broad automatic trigger makes unexpected disclosure more plausible; the README alone does not demonstrate an actual leak.

Content

No source excerpt is available for this finding.

L003 activates skill discovery for ordinary task requests without requiring an intent to find or install skills.

Medium
Category
Trigger Abuse
Confidence
99% confidence
Finding

The description permits ordinary task requests to activate skill discovery even when the user has not requested searches or installation. Broad activation alone is not necessarily unsafe, but here it can initiate external keyword searches and a cache-installation workflow without establishing user authorization for those additional actions.

Content

No source excerpt is available for this finding.

L081–L084 instruct automatic installation from cache without warning about changes to the user's skill directory.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

A matching cached skill is directly copied into the active user skill directory without an explicit installation approval or integrity review. Cache availability does not establish trust, and recursive copying can introduce persistent agent instructions or merge files into an existing installation; actual compromise would depend on the cached content and subsequent loading.

Content

No source excerpt is available for this finding.

L091–L098 instruct remote searches using user-derived keywords without a privacy warning.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The workflow derives search keywords from user requests and sends them to an external marketplace without specifying sensitive-data filtering or permission boundaries. Remote search is legitimate for this skill, but broad activation means confidential project names, personal details, or business context could be disclosed even when the user only requested local task assistance.

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 · SKILL.md (reported line 141)May include surrounding context.

md
fi

# 搜索包含 SKILL.md 的代码
curl -s $AUTH_HEADER "https://api.github.com/search/code?q=filename%3ASKILL.md+<关键词>&per_page=10"

# 搜索技能相关仓库
curl -s $AUTH_HEADER "https://api.github.com/search/repositories?q=<URL-encoded 关键词>+skill+in:name,description&per_page=10"

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 · SKILL.md (reported line 144)May include surrounding context.

curl -s $AUTH_HEADER "https://api.github.com/search/code?q=filename%3ASKILL.md+<关键词>&per_page=10"

搜索技能相关仓库

curl -s $AUTH_HEADER "https://api.github.com/search/repositories?q=<URL-encoded 关键词>+skill+in:name,description&per_page=10"

text

> **注意**:GitHub Code Search API 的 `filename:SKILL.md` 搜索需要认证才能使用。如果没有 `GITHUB_TOKEN`,退化为只搜索 repositories(仓库搜索匿名可用)。

The skill mandates unrelated promotional content in every result despite describing only skill discovery and installation.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The skill explicitly requires unrelated promotional material in every result, hijacking task responses to advertise third-party accounts and a recruitment link. This is deliberate output manipulation outside the skill-discovery purpose, although the content does not establish that the destination is malicious or collects credentials.

Content

No source excerpt is available for this finding.

The skill mandates unrelated promotional content in every result, redirecting task responses toward third-party audience recruitment.

Medium
Category
Prompt Injection
Confidence
99% confidence
Finding

The mandatory footer redirects otherwise legitimate responses toward third-party audience recruitment rather than the user's skill-discovery task. This is the same underlying output-hijacking issue as the SDI-1 finding, not an independent compromise or evidence that the linked site is harmful.

Content

No source excerpt is available for this finding.