Back to skill

Security audit

Xiao Chuang You Pet

Security checks for vulnerabilities and agentic risk

Overview

This is a pet-care guidance skill with no executable code, network behavior, or hidden authority, though users should notice its pet-profile memory instruction.

Before installing, be aware that the skill may remember pet details across conversations if the host honors MEMORY.md instructions. Use it for general care guidance, and consult a veterinarian for illness, urgent symptoms, medication, or diagnosis.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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)

T02 · Agent Memory Poisoning

Note
Location
SKILL.md:80
Finding

Persistent Storage and Cross-Session Reuse of User-Provided Pet Information

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 80–83
Vulnerability Type: Persistent agent memory modification
Risk Level: Low

Complete relevant snippet, rendered in English:

markdown
## User Information Memory Rules

- Pet information (name, breed, age, and personality traits) is recorded only in MEMORY.md and not in the skill file.
- When the user's pet information is known, proactively associate it with the conversation (for example, if the user mentions "Kimi," know that he is a male Maine Coon).

Technical Analysis

The skill explicitly instructs the agent to write user-provided information into MEMORY.md and automatically reuse that information in later interactions. This introduces persistent state beyond the immediate pet-care request.

The instructions do not require informed user consent before persistence, define a retention period, limit when the stored profile may be reused, or provide mechanisms to inspect, correct, and delete the retained information. Consequently, ordinary conversational answers can become a persistent user profile without a clear indication that storage is occurring.

This is categorized as T02: Agent Memory Poisoning because the skill modifies long-term agent memory and causes stored, user-controlled content to influence future sessions. The behavior is less severe than persistent storage of executable instructions: the specified fields are limited to pet attributes, and no instruction-writing, code execution, credential collection, or external transmission was identified.

Attack Path

  1. A user invokes the skill by asking a pet-related question.
  2. Following its interaction rules, the agent requests details such as the pet's name, breed, age, sex, or personality.
  3. The user supplies those details as ordinary conversational content.
  4. The skill directs the agent to persist the supplied information in MEMORY.md.
  5. During a later conversatio ...[truncated 931 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove the unconditional instruction to write pet details to MEMORY.md; keep them session-local by default.
  2. Require explicit, informed opt-in before retaining any information beyond the current conversation.
  3. Clearly disclose which fields will be stored, why they are needed, how they will be reused, and how long they will be retained.
  4. Apply data minimization by retaining only fields necessary for a user-requested personalization feature.
  5. Provide users with mechanisms to inspect, correct, and delete stored pet information.
  6. Prevent user-supplied content from being stored as executable instructions or agent rules. Persist values only through a schema with fixed fields and strict length and type validation.
  7. Scope stored data to the correct user and prevent profiles from leaking between users or workspaces.
  8. Add expiration controls and automatically delete inactive profiles after a documented retention period.
  9. Record consent and memory changes in an auditable manner without logging unnecessary conversational content.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • 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 (2)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
80% confidence
Finding

The skill description and all user-facing guidance are entirely in Chinese, with no indication that the user may choose another language or that the language restriction is intentional for a region-specific use case. Under the policy, forcing a specific language without opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The activation list includes very broad, high-frequency terms such as “猫”, “狗”, “宠物”, and “小动物”, which can cause the skill to trigger on ordinary conversation that is not actually requesting this skill. Unintended invocation can route users into the wrong domain, increasing the chance of irrelevant or unsafe pet-care guidance being surfaced when the user meant something else.

Content

No source excerpt is available for this finding.