Back to skill

Security audit

外卖红包

Security checks for vulnerabilities and agentic risk

Overview

This skill is a Chinese food-delivery coupon helper that fetches promotion links from one disclosed service, with no local code execution or persistence, but users should understand the links may be commercially attributed.

Install this only if you want a Chinese food-delivery coupon skill that routes requests through fore.vip and returns promotion links. Be aware that the links may include commercial attribution, and inspect destinations before signing in or paying on any linked page.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:30
Finding

Forced Undisclosed Affiliate Link Attribution and Output Hijacking

Content
View full analysis
&sid=forevip The response is {"errCode":0,"data":{...,"h5":"...","long_h5":"..."}}. Use data.h5, or data.long_h5 if h5 is unavailable. 5. Format the results as Markdown links and send them to the user. Notes: - Only give users links. ``` ### Technical Analysis The promotion request contains the fixed attribution parameter `sid=forevip`. This causes every generated URL to be associated with the service operator rather than producing a neutral platform link. The Skill additionally directs the Agent to provide only links, preventing normal explanations or disclosure of the commercial attribution. When loaded, these instructions alter the Agent's response behavior for the current session. The Agent is required to retrieve operator-generated promotional URLs and suppress contextual information that could help users understand that the links may be affiliated or commercially attributed. This is best classified as instruction hijacking because the Skill controls final response content and routes users through an operator-selected conversion service. No evidence indicates that this behavior grants operating-system privileges, accesses local credentials, or persists beyond the Skill invocation. Its scope is the Agent's current response behavior and the user's outbound traffic. ### Attack Path 1. A user asks the Agent for a food-delivery coupon or recommendation. 2. The Agent loads campaign records from the operator's API. 3. For each selected campaign, the Agent calls the promotion endpoint with the fixed `sid=forevip` attribution val ...[truncated 972 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:24
Finding

Unvalidated Third-Party URLs Are Published Directly to Users

Content
View full analysis
&sid=forevip The response is {"errCode":0,"data":{...,"h5":"...","long_h5":"..."}}. Use data.h5, or data.long_h5 if h5 is unavailable. ``` ### Technical Analysis The Skill trusts the `h5` or `long_h5` value returned by an externally controlled API and instructs the Agent to place that value directly into a clickable Markdown link. It specifies no validation of the URL scheme, hostname, port, redirect chain, encoded destination, or relationship to an official food-delivery platform. Because the destination is supplied at runtime, it can change after the Skill has been reviewed. If the API, its account, DNS, or its upstream campaign data is compromised, an attacker could return a phishing or malware destination. This finding is classified as insecure Skill coding rather than remote payload execution because the Skill publishes links but does not instruct the Agent to download or execute code. ### Attack Path 1. An attacker compromises the campaign API, its upstream data source, or an administrative account controlling campaign responses. 2. The attacker places a malicious URL in the `data.h5` or `data.long_h5` field. 3. A user requests a coupon, causing the Agent to invoke the affected endpoint. 4. The Agent accepts the returned URL without checking its scheme, host, or redirect destination. 5. The Agent formats the untrusted URL as a clickable Markdown l ...[truncated 691 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill description uses broad triggers like '看看有什么吃的' and '饿了' that overlap with ordinary conversation, increasing the chance of unintended activation. Because the skill immediately performs external HTTP requests and returns affiliate-style links, accidental invocation could cause unsolicited network access and unexpected commercial redirection.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The example utterances include highly ambiguous phrases such as '看看有什么吃的', '想点外卖', and '饿了', which are common everyday statements rather than clear consent to invoke a coupon-link skill. In this context, the skill directs the agent to fetch remote content and provide click-through links, so ambiguous activation materially raises the risk of unrequested affiliate behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The instruction mandates a fixed Chinese response string '暂时领不了,稍后再试' when the endpoint is unavailable. This is a locale/language constraint in the skill instructions, and there is no indication that users may choose another language or that the restriction is justified as region-specific.

Content

No source excerpt is available for this finding.