Back to skill

Security audit

Persona Forge

Security checks for vulnerabilities and agentic risk

Overview

PersonaForge is a coherent AI companion skill, but it handles sensitive chat history, persistent memory, scheduled proactive messages, and messaging-channel setup with insufficiently tight consent and data-flow boundaries.

Review before installing if you plan to upload real chat logs, screenshots, images, or third-party conversations. Use summaries instead of raw transcripts where possible, avoid sensitive or non-consensual content, confirm any channel/API setup before enabling it, and check or remove created state files and cron jobs if you stop using the 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
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (9)

Context Leakage

High
Category
Data Exfiltration
Confidence
94% confidence
Finding

The skill explicitly encourages uploading chat logs and using them to generate persistent companion behavior and memory. Chat logs commonly contain sensitive personal data and third-party conversations; combined with long-term storage and personality analysis, this creates a real risk of over-collection, privacy leakage, and unintended retention of highly sensitive context.

Content

Scanner excerpt · SKILL.md (reported line 3)May include surrounding context.

md
---
name: personaforge
description: "情感陪伴AI | AI Companion. 上传素材自动生成角色,按性格主动聊天,记住用户特点,随时间自然成长。Upload chat logs, images, or descriptions — auto-generate characters that chat based on personality, remember you, and grow over time."
---

# PersonaForge — 情感陪伴AI · AI Companion

Context Leakage

High
Category
Data Exfiltration
Confidence
91% confidence
Finding

The document acknowledges third-party conversations may be uploaded and places the burden on the user to obtain consent, but that alone is not a sufficient privacy control. The skill still facilitates ingestion of third-party data into persistent memory/state, which can expose non-users' private information without robust technical safeguards.

Content

Scanner excerpt · SKILL.md (reported line 204)May include surrounding context.

md
**第三方数据提醒 Third-party data:**
> ⚠️ 如果你上传的聊天记录包含他人的对话,请确保你已获得对方同意
> ⚠️ If you upload chat logs containing third-party conversations, ensure you have obtained their consent

### 如何停止使用 · How to Stop

Context Leakage

High
Category
Data Exfiltration
Confidence
95% confidence
Finding

This onboarding flow directly prompts users to upload chat logs for analysis and generation, increasing the likelihood that sensitive or regulated data will be provided during initial setup. Because the skill also advertises memory and long-lived state, the prompt materially raises the chance of privacy harm through persistent storage, profiling, and reuse of intimate conversational context.

Content

Scanner excerpt · SKILL.md (reported line 241)May include surrounding context.

md
> - 或者描述你想要的性格:『粘人的女朋友』『高冷的学长』
>   Or describe the personality: "a clingy girlfriend", "an aloof senior"
> - 或者上传聊天记录、人物设定,我来帮你分析生成
>   Or upload chat logs or character notes — I'll analyze and generate
>
> 或者输入 `help` 查看所有功能
> Or type `help` to see all commands」

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill invites users to type help to view all commands, but the document does not define a bounded invocation scope or distinguish this from ordinary conversational use. In a chat-based skill, such a generic trigger can overlap with common user speech and may cause unintended activation or routing.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The entire template is written in Chinese and all example fields, labels, and expected outputs are framed exclusively in Chinese, with no indication that other languages are supported. This creates a natural-language locale constraint without user opt-in, which matches the policy category for forced language or locale.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The guide states that 'all data stays local' while also disclosing behaviors that can transmit user-derived data externally, including web searches triggered by character names and optional third-party API integrations for image, TTS, and STT features. In a skill centered on sensitive chat logs, images, and persistent memory, this is a materially misleading privacy claim that can cause users to expose personal or third-party data under false assumptions.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The quick-start examples authorize broad actions from minimal user input, such as analyzing uploaded chat logs, generating persistent profiles, configuring channels, creating cron tasks, and starting proactive messaging. Without explicit trigger boundaries, confirmation steps, or consent gates, a user can unintentionally invoke privacy-sensitive processing and autonomous behavior from casual prompts.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
78% confidence
Finding

The file hardcodes bilingual prompts and several Chinese-led examples, but it does not explicitly ask the user which language they prefer. This can be a locale-policy concern if organizational policy requires user opt-in before enforcing or defaulting to a language or locale.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
81% confidence
Finding

The guide is predominantly written in Chinese and all user-facing examples, behavior descriptions, and reminder content assume Chinese-language interaction, but it does not state that language is optional or user-selectable. This can conflict with language/locale policy expectations when no opt-in or alternative locale handling is provided.

Content

No source excerpt is available for this finding.