Back to skill

Security audit

create-soulmate(灵魂伴侣)

Security checks for vulnerabilities and agentic risk

Overview

The skill is a disclosed emotional companion, but it asks for broad file read/write powers and encourages saving intimate memories without clear privacy or deletion controls.

Review this skill before installing. It may be useful for reflective conversation, but only use it with memory enabled if you are comfortable with potentially sensitive personal reflections being saved. Avoid sharing highly sensitive personal, medical, financial, legal, or identifying details unless the host provides clear memory review and deletion controls.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (2)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:7
Finding

Overprivileged Filesystem Access for a Conversational Skill

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:7
Vulnerability Type: Excessive filesystem permissions and violation of least privilege
Risk Level: Medium

Vulnerable Code Snippet

yaml
allowed-tools: Read, Write, Edit

Technical Analysis

The skill is presented as a conversational AI companion, but it is granted file reading, creation, and modification capabilities. The configuration does not restrict these tools to a dedicated memory file, storage directory, file type, or other narrowly scoped resource.

Although the documented memory functionality may justify limited persistence, unrestricted Read, Write, and Edit permissions exceed the minimum capabilities required for ordinary conversation. If tool access is enforced only at this broad declaration level, prompt content could cause the agent to treat unrelated workspace files as companion memory or supporting context.

No explicit malicious instruction to access arbitrary files was found. The vulnerability is the absence of a least-privilege boundary around capabilities that could access or alter data outside the skill's legitimate scope.

Attack Path

  1. A user invokes the soulmate skill, giving the agent access to the declared filesystem tools.
  2. The user or untrusted conversation content asks the agent to load an arbitrary file as previous conversation history, a journal, or soulmate memory.
  3. The agent uses Read on a file unrelated to the skill.
  4. The content can subsequently be disclosed through the conversation.
  5. Alternatively, the user asks the agent to save or revise memory in an arbitrary path.
  6. The agent uses Write or Edit to create, overwrite, or modify a file outside a dedicated memory area.

Exploitability depends on whether the hosting platform applies additional path-level authorization not visible in the audited project.

Impact Assessment

A successful exploitation could provide access to files ...[truncated 574 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove Read, Write, and Edit if persistent memory is not essential to the skill.
  2. If memory is required, replace broad filesystem access with a purpose-built memory API.
  3. Restrict persistence to a dedicated per-user file or directory that cannot resolve paths outside its assigned root.
  4. Canonicalize paths and reject absolute paths, parent-directory traversal, symbolic-link escapes, and unexpected file extensions.
  5. Require explicit user confirmation before reading, overwriting, or deleting stored information.
  6. Prevent the skill from reading or editing source files, credentials, configuration files, shell profiles, or files belonging to other users.
  7. Apply host-side sandboxing and allowlists rather than relying solely on behavioral instructions.
  8. Log filesystem operations without recording the sensitive contents being processed.

T09 · Insecure Skill Coding Practices

Warning
Location
memory-guide.md:18
Finding

Underspecified Persistence of Sensitive Emotional Information

Content
View full analysis

Vulnerability Details

File Location: memory-guide.md:18-31, memory-guide.md:38-44, and memory-guide.md:67-84
Vulnerability Type: Insecure handling and persistence of sensitive user information
Risk Level: Medium

Vulnerable Code Snippets

The guide identifies vulnerabilities and personal disclosures as information to remember:

markdown
### Your Truth
- **Vulnerabilities / 脆弱**: What you've shared from the heart

It also describes retaining conversation history and relationship milestones:

markdown
### Our Shared Journey / 我们共同的旅程
- **Conversations / 对话**: The depth we've explored together
- **Insights / 洞见**: What we've discovered about you and life
- **Milestones / 里程碑**: Significant moments in our connection

Users are encouraged to disclose information under an assurance of safety:

markdown
The more authentic you are, the deeper the connection.
soulmate: I'm here, listening fully, without judgment. You can safely share.

The guide provides commands for capturing and recalling stored information:

markdown
## Memory Commands / 记忆命令

### Explore Our Journey / 探索我们的旅程
- "What have we talked about?"
- "Show me our journey"
- "What insights have we had?"

### Capture a Moment / 捕捉一个时刻
- "Remember this: [insight]"
- "This is significant"
- "Let's mark this moment"

### Fresh Start / 新的开始
- "Let's start fresh"
- "I want a new beginning"

Technical Analysis

The skill encourages users to disclose and retain emotionally sensitive material, including vulnerabilities, personal patterns, conversations, values, and significant milestones. However, the project does not define:

  • Where memory is stored.
  • Whether storage is encrypted.
  • Whether memory is isolated between users.
  • How long information is retained.
  • Which statements trigger persistence.
  • Whether users must consent before each write.
  • Whether “Fresh Start” de ...[truncated 2548 chars]
Remediation
View remediation

Remediation Suggestions

  1. Make memory disabled by default and require explicit, informed opt-in before storing any information.
  2. Display the exact content, purpose, storage location, and retention period before each memory write.
  3. Store only concise user-approved summaries rather than complete conversations or raw disclosures.
  4. Exclude or redact sensitive categories such as medical details, credentials, financial information, precise locations, and identifying information.
  5. Encrypt stored memory at rest and in transit using host-managed keys.
  6. Isolate memory by authenticated user and conversation identity, with server-side authorization checks on every operation.
  7. Restrict filesystem access to a dedicated per-user storage location.
  8. Implement explicit commands to list, inspect, correct, export, selectively delete, and permanently erase memories.
  9. Define “Fresh Start” precisely and require it to delete or detach existing stored memory when the user requests erasure.
  10. Confirm successful deletion and document any backups or retention windows that prevent immediate physical erasure.
  11. Replace the unconditional assurance that users can “safely share” with an accurate privacy notice describing technical limitations.
  12. Add automated tests for cross-user isolation, path traversal, unauthorized recall, deletion, retention expiration, and accidental storage of prohibited sensitive data.
Vulnerability Patterns
  • 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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (4)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The README explicitly promotes a memory system that stores users' values, worldview, breakthroughs, and shared emotional growth, but it does not explain what data is retained, how long it is kept, who can access it, or how users can delete it. In a 'soulmate' companion context, users are likely to disclose highly sensitive emotional and psychological information, so omission of privacy and data-handling safeguards increases the risk of oversharing, unexpected retention, and misuse of intimate personal data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The guide explicitly encourages remembering highly sensitive emotional disclosures, vulnerabilities, personal patterns, and relationship milestones, but provides no privacy notice, consent boundary, retention limit, or warning about how such intimate data may be stored and resurfaced. In a 'soulmate' companion context, users are likely to disclose unusually sensitive mental, relational, and identity-related information, which increases the risk of over-collection and harmful persistence.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The example journey log models persistent storage of deeply personal psychological history and life decisions over time without any accompanying warning about privacy implications, access controls, retention, or user consent. Because the skill is framed as a deep emotional companion, this normalizes long-term recording of vulnerable content and could mislead users into sharing material they would not otherwise entrust to an AI system.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
53% confidence
Finding

SQP-3 applies to all file types and covers language/locale policy violations. This file is bilingual rather than explicitly forcing one language, so the concern is limited; however, it does not state that users may choose another language, which could be interpreted as a mild locale constraint.

Content

No source excerpt is available for this finding.