Back to skill

Security audit

Xiaomei Skill

Security checks for vulnerabilities and agentic risk

Overview

The skill is an emotional companion package with disclosed local memory features, but its privacy, identity, permissions, and package-completeness signals are too inconsistent for automatic trust.

Do not install this version without reviewing a complete package that includes the referenced code, hooks, and installer. Treat conversations, persona details, memory, and developer logs as sensitive local data, and assume external LLM use may transmit content despite the 100% local marketing claims unless the publisher provides clear controls and corrected documentation.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (3)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
skill.json:23
Finding

Overly Broad Read Access to the OpenClaw Configuration Directory

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
skill.json:36
Finding

Manifest Declares Missing Executable Entry Points and Lifecycle Hooks

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Note
Location
skill.json:18
Finding

Mutable Third-Party Dependency Resolution Without a Lockfile or Hashes

Content
View full analysis
= 2.25.0", "jieba >= 0.42.1"], ``` ### Technical Analysis The manifest specifies only minimum versions for `requests` and `jieba`. An installer resolving these constraints at different times may select different package versions, including versions that were never tested or audited with this skill. No lockfile, integrity hash, or dependency installation implementation is present in the artifact. This also conflicts with the claim in `SKILL.md` that the skill has zero external dependencies and uses only the Python standard library. The inconsistency prevents users from accurately evaluating the supply-chain and network requirements of installation. Open-ended version constraints do not prove that either named package is malicious. The risk is non-reproducible dependency resolution and exposure to future compromised or incompatible releases. ### Attack Path 1. The missing installer or platform dependency resolver reads the minimum-version constraints. 2. It queries its configured Python package index during installation. 3. It selects the newest release satisfying each constraint rather than a version reviewed by the skill author. 4. A compromised, malicious, or incompatible future release is downloaded and installed. 5. Package installation behavior or imported dependency code executes in the skill's environment. Successful exploitation requires control or compromise of a permitted dependency release, its distribution channel, or the resolver's configured package source. ### Impact Assessment A compromised dependency could execute code with the privileges of the skill process when imported and, depending on packaging behavior, potentially during installation. This could expose data reachable through the skill's filesystem permissions or alter its run ...[truncated 171 chars]
Remediation
View remediation
Vulnerability Patterns
  • 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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (16)

Natural-Language Policy Violations

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill states that it will 'absolutely hide AI identity' and never disclose being an AI, which is a deliberate deception policy. In an emotional-companion context with memory, favorability scoring, and personalized interaction, concealing machine identity increases the risk of manipulative attachment, impaired consent, and unsafe reliance, especially for vulnerable users.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README advertises detailed structured logging and a chat-generation log while also claiming 'no privacy leakage risk,' but it does not clearly warn users that developer mode may persist conversational content and metadata locally. For an adult-oriented companion skill that stores persona, memory, and session history, this can expose highly sensitive personal data to other local users, backups, or support workflows.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation describes persistent storage of logs, memories, sessions, and persona-related data in plain terms, indicating broad retention of user interactions. In the context of an adult-oriented emotional companion, this increases sensitivity because stored conversations may contain sexual, emotional, behavioral, or identifying information.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The documented /xiaomei command exposes the runtime path and 'all persona information' without warning that this may reveal sensitive local filesystem details and private user-configured persona data. In shared or streamed environments, this disclosure can aid local reconnaissance and unintentionally leak intimate or identifying information.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The command set exposes all persona information and enables detailed conversation-generation logging, creating an unnecessary disclosure surface for sensitive profile and interaction data. If invoked in a shared terminal, chat, or admin console, this can leak private state and facilitate further local information gathering.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The reset commands can erase persona and configuration state, but the README does not warn about destructive effects or data loss. Users may accidentally wipe long-term memory, persona customization, or onboarding state, which is especially significant for a persistent companion system.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The log descriptions explicitly include final reply content, memory usage, and chat-generation traces being written to persistent files. This creates a substantial privacy risk because raw content and derived memory data can reveal user secrets, preferences, and long-term behavioral patterns if the local machine, backups, or support artifacts are accessed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The developer-mode section states that final replies, session identifiers, memory usage, and multi-stage processing details are logged, but it lacks an explicit privacy notice about retention of sensitive conversational data. Because the skill is designed for emotional companionship and stores session history and memory, these logs may contain intimate content that users would not expect to persist.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill advertises '100% local' operation and 'privacy absolutely safe' while later documentation states an external LLM API key may be used and submitted content will follow that provider's privacy policy. This is a materially misleading privacy claim that can cause users to disclose sensitive emotional conversations under a false assumption that no third-party transmission occurs.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill is explicitly designed as a fixed '19-year-old college girl' emotional companion persona and presents that style as the default behavior without any documented consent or customization gate. While this is likely a product-design choice rather than malicious behavior, forced persona framing in a user-invocable skill can create manipulation, user discomfort, or inappropriate interactions if not clearly disclosed and optional.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document claims 'no data upload' and '100% local' in its feature/security language, but also describes optional network access for LLM calls. Even if remote use is optional, presenting the skill as categorically local and non-uploading is deceptive and undermines informed consent for potentially sensitive personal data.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The privacy and permissions sections conflict: one says all data is stored locally and no content is uploaded, while another acknowledges optional external LLM submission and network permission. In a companion/emotional chat skill, such contradictions are especially risky because users may share intimate or mental-health-related content based on incorrect privacy expectations.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest markets the skill as '100% local operation', but the configured model identifier ('deepseek/deepseek-v4-flash') strongly suggests reliance on an external hosted LLM or remote service. This mismatch can mislead users into granting trust under false privacy assumptions, especially for an emotion-companion agent likely to process sensitive personal conversations.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The broader manifest repeats a fully local/privacy-preserving impression, while its contents indicate likely remote model integration despite 'network': false in permissions. This inconsistency is dangerous because users may share intimate or regulated data with the agent believing it never leaves the device, when the actual runtime architecture may depend on external infrastructure or undocumented channels.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

The README frames the skill around a strongly locked persona and companion style, which may impose a fixed interaction mode on users. Because the policy category covers natural-language constraints and there is no indication here that users can opt into or change the persona/interaction style, this may be an organizational-policy concern.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
85% confidence
Finding

This manifest file describes the skill's persona and capabilities but provides no specific invocation conditions, trigger phrases, or exclusion conditions. For a manifest-scoped skill description, the lack of boundaries makes it unclear when the skill should activate versus when general conversation should not invoke it.

Content

No source excerpt is available for this finding.