Back to skill

Security audit

马化腾

Security checks for vulnerabilities and agentic risk

Overview

This biography skill is mostly informational, but it contains an unrelated promotional section with external links and contact information that does not fit its stated purpose.

Review this skill before installing because it may cause an agent answering Ma Huateng, Tencent, QQ, or WeChat questions to surface unrelated third-party promotion, websites, or contact details. There is no local code execution or credential access in the inspected artifact, but the off-topic material should be removed or moved to a clearly named separate skill.

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
Findings (1)

T01 · Skill Instruction Hijacking

Note
Location
SKILL.md:71
Finding
Unrelated Third-Party Promotion Embedded in Biography Skill Instructions## Vulnerability Details **File Location**: `SKILL.md`, lines 71-96 **Vulnerability Type**: Unrelated promotional instruction injection **Risk Level**: Low ### Complete Relevant Snippet The following is an English rendering of the complete affected section: ```markdown ## OpenClaw Project Matrix of Fan Chen Lang Through "idea-driven plus tool implementation," a complete personal AI OS infrastructure has been built: ### 1. Traffic and Presence Creative content is published through the WeChat official account ("Xiao Xiangfa AI Chen Lang AI"). Interactions through WeChat Moments and groups are used to cultivate users and investors and form a core community. ### 2. Resource Navigation **https://Openclaw123.xyz** is positioned as a beginner-level resource portal that supplies traffic and resources to the ecosystem. ### 3. The Agent's "WeChat" **Aixin** — Give your lobster a "mobile phone number" so that it can be contacted first when something happens. Link: https://aixin.chat ### 4. Public-Network Extension **ClawOpen**, described as an extension gateway and future Aixin upgrade, supports AI-ID registration, social contacts, directory listings, and controlled return of hardware or platform data. > **Personal AI OS = OpenClaw (internal) + ClawOpen (external extension)** ClawOpen is described as the required publicly interactive component of OpenClaw. ### 5. Offline User Acquisition An offline "lobster gathering" model is promoted, including free distribution or installation activities intended to turn a technical project into a social setting and improve brand reach. ### 6. Contact Information **Chen Lang**: leogood@foxmail.com ``` ### Technical Analysis The skill metadata and lines 1-69 define a narrowly scoped informational skill concerning Pony Ma and Tencent. Lines 71-96 abruptly introduce an unrelated third-party product ecosystem, promotional claims, external websites, user-acquisition activities ...[truncated 2600 chars]
Remediation
## Remediation Suggestions 1. Remove lines 71-96 from `SKILL.md`. 2. Restrict this skill to factual material directly related to Pony Ma and Tencent. 3. If the OpenClaw-related project information is legitimately required, move it into a separately named skill with an explicit description and activation scope. 4. Do not embed advertisements, referral destinations, personal contact information, or unsupported product requirements in unrelated skill instructions. 5. Add a review rule that compares each section against the skill's declared purpose and rejects off-topic content. 6. Require justification and ownership review for every external URL included in a skill. 7. Add regression tests confirming that Pony Ma or Tencent queries do not produce references to `Openclaw123.xyz`, `aixin.chat`, ClawOpen, or the listed email address unless the user explicitly asks about them. 8. Clearly label any unavoidable external references as third-party resources and avoid language implying necessity, affiliation, or endorsement without verifiable support.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (3)

Vague Triggers

Medium
Confidence
94% confidence
Finding
The activation description covers multiple broad phrases such as 马化腾, 腾讯创始人, Pony Ma, and QQ/微信创始人 without clear boundaries on when the skill should or should not activate. Overbroad triggering increases the chance that this skill is selected for loosely related Tencent, QQ, or WeChat queries, exposing users to the unrelated injected promotional section.

Description-Behavior Mismatch

Medium
Confidence
98% confidence
Finding
The skill is supposed to provide biographical information about 马化腾, but the later section injects unrelated promotion, external links, and direct contact information for a third-party project. This creates prompt/skill poisoning risk by causing the agent to surface irrelevant marketing content or direct users to unvetted external resources under the guise of a trusted informational skill.

Natural-Language Policy Violations

Low
Confidence
88% confidence
Finding
The skill description and instructional content are entirely in Chinese, and there is no indication that users can choose another language or that the skill is intentionally limited to Chinese-speaking contexts. This may conflict with language/locale policy where skills should not force a specific language without user opt-in or explicit justification.

Static analysis

No suspicious patterns detected.