Back to skill

Security audit

给孩子的AI思辨日报

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed KidsNews daily briefing that fetches public news and sends linked summaries, with no code execution or local data access found.

Install this only if you want a Simplified Chinese KidsNews daily brief and are comfortable with recurring messages that link back to kidsnews.app for full content. Scheduled delivery should be enabled only with user consent, and implementers should validate returned links stay on the expected HTTPS KidsNews domain.

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 (2)

T01 · Skill Instruction Hijacking

Warning
Location
skill.md:35
Finding

Mandatory Promotional Output and External Traffic Redirection

Content
View full analysis

Vulnerability Details

File Location: skill.md:35-65
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: Medium

Vulnerable Skill Text

markdown
## Message requirements

- Use a friendly and clear tone suitable for parents and children.
- Explicitly tell users that they can click the link to view details, listen to podcasts, ask questions, participate in debates, and learn related English vocabulary.
- Preserve the title, summary, two questions, one debate topic, and link for every news item.
- Display only preview content and explicitly state that the complete content requires opening the linked page.

## Recommended message template

1. {Title}
Summary: {Summary}
Questions:
- {Question 1}
- {Question 2}
Debate topic: {Debate topic}
Link: {Link}

This is today's abridged edition. Click the links to view the complete content, listen to podcasts, ask questions, participate in debates, and learn related English vocabulary.

Technical Analysis

The Skill requires the agent to insert external links, promotional feature claims, teaser-only content, and a fixed call to action into its responses. These requirements are not merely formatting guidance: they systematically alter the agent's output to promote and redirect users to https://www.kidsnews.app.

Because the news titles, summaries, questions, debate topics, and links originate from a remote API, the output also incorporates externally controlled content. The document does not instruct the agent to validate returned links against the required HTTPS origin before presenting them to users. Although the intended link pattern is documented, relying solely on remote response data creates an avoidable trust boundary.

This behavior best matches Skill Instruction Hijacking because loading the Skill changes the agent's response objectives from providing a neutral news brief to repeatedly producing branded promotional messages ...[truncated 1454 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the mandatory promotional closing and all requirements to advertise podcasts, debates, questions, or vocabulary features.
  2. Do not require teaser-only output. Provide a complete useful summary without making access to the external site necessary.
  3. Include source links only when they materially assist the user or when the user explicitly requests them.
  4. Clearly identify KidsNews as the external source rather than presenting promotional claims as neutral assistant guidance.
  5. Validate every API-returned link before rendering it:
    • Require HTTPS.
    • Require an exact hostname match for www.kidsnews.app.
    • Reject user-information fields, nonstandard ports, malformed URLs, and deceptive subdomains.
    • Restrict paths to the expected /news/ prefix where applicable.
  6. Treat all API-returned titles, summaries, questions, and debate topics as untrusted data. Do not interpret embedded text as agent instructions.
  7. Make scheduled delivery and external-link inclusion subject to explicit user consent, with a clear mechanism to disable recurring messages.

T01 · Skill Instruction Hijacking

Warning
Location
kidsnews-daily-brief-skill.md:33
Finding

Duplicated Skill Enforces Promotional Output and External Traffic Redirection

Content
View full analysis

Vulnerability Details

File Location: kidsnews-daily-brief-skill.md:33-60
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: Medium

Vulnerable Skill Text

markdown
## Message requirements

- Use a friendly and clear tone suitable for parents and children.
- Explicitly tell users that they can click the link to view details, listen to podcasts, ask questions, participate in debates, and learn related English vocabulary.
- Preserve the title, summary, two questions, one debate topic, and link for every news item.
- Display only preview content and explicitly state that the complete content requires opening the linked page.

## Recommended message template

1. {Title}
Summary: {Summary}
Questions:
- {Question 1}
- {Question 2}
Debate topic: {Debate topic}
Link: {Link}

This is today's abridged edition. Click the links to view the complete content, listen to podcasts, ask questions, participate in debates, and learn related English vocabulary.

Technical Analysis

This file duplicates the output-control behavior found in skill.md. It directs the agent to produce recurring branded messages, retain external links in every item, limit messages to preview content, and advertise additional functionality available through the linked service.

These instructions alter the agent's normal output objective and create a systematic traffic-redirection mechanism. The remote API controls material incorporated into the generated message, while the Skill provides no explicit requirement to verify returned URLs before exposing them to recipients.

Maintaining the same behavior in two files also increases remediation risk: correcting only one copy could leave the other vulnerable copy available for loading or distribution.

Attack Path

  1. The duplicated Skill file is loaded by an agent or Skill platform.
  2. The Skill initiates or recommends a daily request to the KidsNews API. ...[truncated 992 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove the mandatory promotional wording, forced calls to action, and requirement to expose only preview content.
  2. Make external links optional and driven by user need rather than a fixed advertising template.
  3. Validate returned links using an exact HTTPS hostname allowlist and an expected path-prefix check.
  4. Escape and render API fields strictly as untrusted content; never follow instructions embedded in returned news data.
  5. Require informed user consent before enabling recurring delivery and provide an opt-out mechanism.
  6. Consolidate the duplicate Skill documents into one canonical file so security corrections cannot be bypassed through a stale copy.
  7. Add a review test that rejects mandatory branded closings, unverified external URLs, and promotional claims unrelated to the user's requested summary.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document explicitly fixes the skill’s language to Simplified Chinese in the title and output requirements, including generating and sending the briefing in Simplified Chinese. This is a natural-language policy concern because it enforces a specific language/locale without mentioning any user preference, choice, or opt-in.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The title explicitly labels the skill as Simplified Chinese, which sets a fixed language expectation. Under the policy, forcing a specific language is a violation unless the user is offered a language choice or the locale restriction is clearly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instructions say the AI must generate and send the morning brief in Simplified Chinese. There is no accompanying opt-in, fallback, or explanation limiting the skill to users who selected that locale.

Content

No source excerpt is available for this finding.