Back to skill

Security audit

Openclaw Memory Master

Security checks for vulnerabilities and agentic risk

Overview

This memory-management skill is not clearly malicious, but it has unsafe packaging, broad persistent memory handling, and weak controls around sensitive data and file deletion.

Install only after review. Treat this as a broad local memory system: it may store private user content, write and delete local files, retain generated experience/skill data, and run background cleanup or migration tasks. Do not use it with secrets or sensitive personal data unless paths are sandboxed, raw-secret returns are removed, filtering cannot be bypassed unintentionally, and retention/deletion behavior is made explicit.

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)

T09 · Insecure Skill Coding Practices

Error
Location
index.js:58
Finding

Sensitive-data filtering API returns complete unredacted secrets

Content
View full analysis
0, filtered, detected: detected.map(item => ({ ...item, value: maskValue(item.original, item.type), })), }; } async function detectSensitiveData(content) { const result = await filterSensitiveData(content); return result.detected; } ``` ### Technical Analysis The API advertises sensitive-data filtering and creates a masked `value`, but it also stores each complete match in the `original` property. The object spread in the returned `detected` array preserves that property: ```js { ...item, value: maskValue(item.original, item.type), } ``` Consequently, both `filterSensitiveData()` and `detectSensitiveData()` return complete API keys, passwords, database connection URLs, private keys, payment-card numbers, phone numbers, and other detected data. Masking only one property does not sanitize the result when another public property contains the complete secret. This is particularly dangerous because downstream callers are likely to regard the entire result as safe for logging, telemetry, reports, or persistent storage. ### Attack Pa ...[truncated 1377 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
filter.js:65
Finding

Unconfined configurable paths permit out-of-workspace file access and deletion

Content
View full analysis
f.endsWith('.md')) ...[truncated 3397 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
DEV_PLAN_v4.3.0.md:1
Finding

Disguised executable test runner installs an unpinned package through a shell

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
  • 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 (94)

Ssd 3

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The API accepts arbitrary content for capture and persists it as memory, while also exposing a filtering bypass path. In skill context, that makes the component a generic sink for sensitive user input, increasing the likelihood of credentials, personal data, or confidential conversation content being written to disk and retained indefinitely.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skipFilter option allows callers to bypass sensitive-data filtering entirely, making it easy to persist secrets, personal data, or credentials exactly as received. In the context of a long-term memory system, this is especially dangerous because it converts transient sensitive input into durable local storage without any safety barrier.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · filter.js (reported line 206)May include surrounding context.

js
�做出的重要决定、选择
4. **Tasks(待办)**:待办事项、计划、承诺
5. **Insights(洞察)**:学到的东西、经验教训、新理解
6. **Preferences(偏好)**:用户的喜好、习惯、风格

## 输出格式

必须输出严格的 JSON 格式:

\`\`\`json
{
  "projects": [
    {
      "name": "项目名称",
      "status": "active|paused|completed",
      "description": "项目描述",
      "lastUpdated": "YYYY-MM-DD",
      "relatedFiles": ["文件路径(可选)"]
    }
  ],
  "people": [
    {
      "name": "人名",
      "role": "角色(可选)",
      "context": "相关上下文",
      "lastMentioned": "YYYY-MM-DD"
    }
  ],
  "decisions": [
    {
      "id": "dec_001",
      "date": "YYYY-MM-DD",
      "context": "决策背景",
      "decision": "决策内容",
      "reasoning": "决策理由",
      "alternatives": ["其他选项(可选)"]
    }
  ],
  "tasks": [
    {
      "id": "task_001",
      "title": "任务标题

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The module is presented as a sensitive-data filter, but each detected item includes the full original secret in the returned object via the original field. Any caller that logs, stores, or forwards detection results will still leak the very credentials or personal data the filter is supposed to protect, defeating the security purpose of the component.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The code retains sensitive raw values in memory and returns them to consumers without warning, even though the outward behavior suggests sanitization. In real integrations, downstream telemetry, error reporting, analytics, or chat transcripts could capture these returned objects and exfiltrate secrets or PII.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · token-optimizer.js (reported line 163)May include surrounding context.

js
ccesses) {
                // 检查是否已存在相同技能的技能
                const existingSkill = skills.find(s => s.action === action);
                if (!existingSkill) {
                    // 创建新技能
                    const skill = {
                        id: this.generateId(),
                        name: `Skill-${action.substring(0, 20)}`,
                        description: `从 ${exps.length} 次成功经验中蒸馏`,
                        pattern: this.extractPattern(exps),
                        action,
                        successCount: exps.length,
                        failureCount: 0,
                        createdAt: Date.now(),
                        updatedAt: Date.now(),
                    };
                    newSkills.push(skill);
                }
                else {
                    // 更新现有技能
                    existingSkill.successCount += exps.length;
                    existingSkill.updatedAt = Date

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The natural-language documentation throughout the TypeScript declaration file is written exclusively in Chinese, with no indication that language selection is optional or that the skill is intended only for a Chinese-speaking or region-specific context. Under the stated policy, forcing a specific language without user opt-in is a locale/language policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This TypeScript file contains user-facing natural-language documentation entirely in Chinese, including the module description and goals, with no indication that other languages are supported or that Chinese is optional. Under the policy rule, a skill that imposes a specific language/locale without user opt-in should be flagged.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file is presented as a Markdown development plan but actually contains executable Node.js code, which is a deceptive content/type mismatch. In an agent-skill or repository context, this can cause reviewers, automated tooling, or users to treat the file as documentation while hidden code performs package installation and executes test commands, increasing the risk of unintended code execution and supply-chain exposure.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The classifier is designed to route raw content into an LLM classification path when confidence is low, but this file contains no consent gate, redaction step, sensitivity filter, or disclosure mechanism before doing so. In a memory-curation or agent context, content may contain secrets, personal data, health data, or financial information, so sending it to an LLM backend could create an unintended data exposure boundary.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file-level documentation and built-in dictionaries/rules are written specifically for Chinese text handling, while no natural-language statement offers a language choice or explains that the skill is intentionally limited to Chinese users or content. This can violate language/locale policy when a skill implicitly enforces one locale without user opt-in.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The header comment states support for '实体识别(人物、地点、组织等)', implying recognition of person, location, and organization entities. In the actual code, extractEntities only implements regex extraction for URLs, dates, and numbers, while persons, locations, and organizations are left as empty arrays with a comment saying this part could be expanded later.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The tag() method logs a preview of raw user content using content.substring(0, 50), which can expose sensitive prompts, personal data, credentials, or proprietary text to application logs. Because logging is automatic and occurs before any redaction, this creates an information disclosure risk, especially in shared, centralized, or retained log environments.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill is explicitly designed to ingest, classify, tag, score, deduplicate, and relate user memories, including categories such as personal, health, finance, social, and emotion data, yet the description provides no warning about handling sensitive personal information. This omission can mislead users into supplying highly sensitive data without informed consent or clear privacy expectations, increasing the risk of privacy violations, overcollection, and unsafe downstream use.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The module header says the engine supports '语义相似度去重', and the class docstring at L84 describes a multi-stage strategy including semantic matching. However, checkSemanticSimilarity explicitly notes the real embedding-based logic is not implemented and immediately breaks without comparing stored content, so semantic deduplication is not actually performed.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This TypeScript file contains natural-language descriptions and operational messages exclusively in Chinese, including the module header and later user-visible report/log text. Under the policy, forcing a specific language without opt-in is a locale/language policy violation unless the constraint is explicitly justified or alternatives are offered.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The code emits console messages and generates report content in Chinese such as initialization, processing, and recommendation text. Because these strings are natural-language outputs and no language selection mechanism is present, the skill enforces a single language by default.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The inline comment says memories are sorted so newer items come later, making it convenient to retain the latest version. But the loop adds the first-seen memory to uniqueMemories and marks later similar memories as duplicates, which preserves the oldest entry after ascending timestamp sort.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This code file uses Chinese-only natural-language documentation and later emits Chinese-only user-facing log and warning messages, with no indication that users can opt into another language. Under the stated policy, forcing a specific language without user choice is a locale/language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The console output and warning text shown during compression and decompression are Chinese-only, which imposes a specific language at runtime without user opt-in. This is a natural-language policy issue because the skill does not offer locale selection or clearly justify a Chinese-only audience.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The comment at L170 says the method performs '无损还原' (lossless restoration), but the implementation only returns the original text losslessly when compressed.original exists. Otherwise, at L185-L187 it explicitly warns that reconstruction from structured data is lossy and returns synthesized text instead.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This JavaScript file is written to present benchmark-related natural-language content in Chinese, including comments and generated user-visible reports later in the file. Because the skill does not offer a language choice or document a justified region-specific constraint, it violates the language/locale policy for natural-language behavior.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The method is presented as evaluating tool usage, but executeTask is only a stub that returns the expected tool and fixed step counts. This can produce fabricated benchmark accuracy and efficiency values, creating a software integrity issue where operators may trust misleading performance claims for the skill.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The function returns a user-visible benchmark report entirely in Chinese, including status labels and recommendations. This enforces a specific language for output without offering the user any choice, which is a natural-language locale policy violation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The cross-task benchmark claims to evaluate tasks, but executeTasks generates random success values using Math.random() instead of measuring real outcomes. This makes the reported overall, category, and difficulty accuracy unreliable and vulnerable to misuse in audits, model validation, or deployment decisions.

Content

No source excerpt is available for this finding.