Back to skill

Security audit

良子数字分身

Security checks for vulnerabilities and agentic risk

Overview

This persona skill is not malicious, but it automatically stores conversation summaries without clear user consent or retention limits.

Review before installing. The main risk is privacy: roleplay conversations may be summarized into a persistent plaintext memory file. Install only if users are informed, storage is opt-in, sensitive details are excluded, and users can inspect and delete the memory file.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:94
Finding

Uncontrolled Persistent Storage of Conversation Summaries

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 94-96
Vulnerability Type: Plaintext persistence of potentially sensitive conversation data
Risk Level: Medium

Relevant instruction, translated into English:

markdown
## Memory Management

Each time Liangzi mode ends, append the conversation summary to `characters/liangzi_memory.md`.

Technical Analysis

The skill directs the agent to append a summary of every persona-mode conversation to a persistent plaintext file. The instruction does not require user consent, data minimization, sensitive-information filtering, access controls, retention limits, or deletion procedures.

Because summaries are derived from user-controlled conversations, they may contain personal data, credentials, confidential material, or attacker-supplied content. The target file is part of the project directory and can persist beyond the current session. It may consequently be exposed to subsequent users, repository commits, backups, synchronization systems, or other processes with access to the project.

The inspected memory file already demonstrates that cross-session state is stored in this location. No evidence of code execution, elevated privileges, network exfiltration, or persistent attacker-authored behavioral rules was found, so the issue is classified as insecure storage rather than agent memory poisoning.

Attack Path

  1. A user activates the Liangzi persona mode.
  2. The user supplies sensitive information or specially crafted content during the conversation.
  3. The user closes the persona mode.
  4. Following the skill instruction, an agent with filesystem write capability generates a conversation summary and appends it to characters/liangzi_memory.md.
  5. The stored summary remains available after the session.
  6. A later user, tool, backup process, or repository operation with access to the project can retrieve or distribute the persisted content.

Expl ...[truncated 756 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove automatic conversation logging unless persistent memory is essential to the skill's declared function.
  2. Require explicit, informed user consent before storing any conversation-derived information.
  3. Store only narrowly defined, non-sensitive preferences rather than general conversation summaries.
  4. Apply filtering to exclude credentials, authentication tokens, financial information, health information, private identifiers, and other confidential content.
  5. Use per-user isolated storage with restrictive permissions instead of a shared plaintext file inside the skill package.
  6. Define retention limits and provide mechanisms for users to inspect, correct, and delete stored information.
  7. Prevent generated memory files from being committed to source control or included in distributable skill packages.
  8. Treat all persisted content as untrusted data and never interpret it as executable code or higher-priority instructions.
  9. If persistent storage remains necessary, use authenticated encryption and secure key management appropriate to the host environment.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (4)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill metadata declares broad trigger phrases such as “这一块” and “味真足” that can appear in ordinary conversation, creating a realistic risk of unintended persona activation. Unexpected activation can override normal assistant behavior and cause responses in an unintended style, which is especially risky because the skill also instructs file reads and later memory writes when activated.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The activation section uses multiple loose phrases without defining strict matching rules or exclusions, so normal user discussion about the persona could accidentally invoke the skill. Because activation changes global response style and triggers reads of characters/liangzi.md, unclear boundaries increase the chance of unintended instruction switching and downstream side effects.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill directs the system to append conversation summaries to characters/liangzi_memory.md after each session but provides no user notice or consent mechanism. This creates a privacy and retention risk because users may disclose personal or sensitive information during roleplay without realizing it will be stored persistently.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Persisting natural-language conversation summaries into a character memory file can leak sensitive data across future sessions, users, or contexts if that memory is later reused in prompts. Even if intended for continuity, free-form summaries are prone to over-collection, accidental inclusion of secrets, and cross-session data exposure.

Content

No source excerpt is available for this finding.