Back to skill

Security audit

蜂兵虾将

Security checks for vulnerabilities and agentic risk

Overview

The skill appears benign overall, with disclosed local memory behavior, though its marketing overstates automation and users should handle stored data carefully.

Install only if you are comfortable with a local-memory tool that stores interaction history, preferences, goals, workflow records, and reports in plain JSON/Markdown files. Do not put secrets or regulated personal data into its memory, keep memory directories private, validate any user-supplied system names if embedding the library, and back up memory before using the documented reset command.

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
dist/core/memory.js:36
Finding
Path Traversal Through Unsanitized Memory and Synchronization Identifiers<![CDATA[ ## Vulnerability Details **File Location**: `dist/core/memory.js`, lines 36-64 and 269-275 **Vulnerability Type**: Path traversal leading to filesystem access outside the intended memory directory **Risk Level**: Medium ### Vulnerable Code ```javascript constructor(skillName = 'ai_system', baseDir = 'memory', config) { this.L0Variables = new Map(); this.L0Context = ''; this.L1Content = ''; this.L2Entries = []; this.L3Data = { worldviews: [], methodologies: [], capabilityModels: [], personalProfile: [] }; this.L4Data = { insights: [], coreValues: [], longTermPredictions: [], inheritableAssets: [] }; this.skillName = skillName; this.baseDir = baseDir; this.config = { L0_MAX_ITEMS: 10, L1_MAX_LINES: 50, L2_MAX_ENTRIES: 200, L3_MAX_ENTRIES: 1000, AUTO_ARCHIVE_THRESHOLD: 0.8, ...config }; this.initialize(); } initialize() { this.createDirectories(); this.loadMemories(); } createDirectories() { const dirs = ['L0_flash', 'L1_working', 'L2_experience', 'L3_knowledge', 'L4_wisdom', 'shared', 'logs']; for (const dir of dirs) { const fullPath = path.join(this.baseDir, this.skillName, dir); if (!fs.existsSync(fullPath)) { fs.mkdirSync(fullPath, { recursive: true }); } } } ``` Additional affected filesystem operations include: ```javascript syncToSystem(targetSystem, entries) { fs.writeFileSync( path.join( this.baseDir, this.skillName, 'shared', `${targetSystem}_sync.json` ), JSON.stringify({ timestamp: new Date().toISOString(), entries }, null, 2), 'utf-8' ); } syncFromSystem(sourceSystem) { const filePath = path.join( this.baseDir, this.skillName, 'shared', `${sourceSystem}_sync.json` ); if (fs.existsSync(filePath)) return JSON.parse(fs.readF ...[truncated 2901 chars]
Remediation
<![CDATA[ ## Remediation Suggestions 1. **Validate all identifiers at runtime** Restrict `skillName`, `targetSystem`, and `sourceSystem` to simple identifiers: ```javascript function validateIdentifier(value, fieldName) { if (typeof value !== 'string' || !/^[A-Za-z0-9_-]+$/.test(value)) { throw new Error(`${fieldName} contains invalid characters`); } return value; } ``` 2. **Establish a fixed trusted storage root** Resolve the configured memory root once and do not permit untrusted callers to supply arbitrary base directories: ```javascript this.baseDir = path.resolve(TRUSTED_MEMORY_ROOT); this.skillName = validateIdentifier(skillName, 'skillName'); ``` 3. **Enforce path containment** Resolve every destination and verify that it remains under the trusted root: ```javascript function resolveWithinRoot(root, ...segments) { const normalizedRoot = path.resolve(root); const candidate = path.resolve(normalizedRoot, ...segments); const relative = path.relative(normalizedRoot, candidate); if ( relative === '..' || relative.startsWith(`..${path.sep}`) || path.isAbsolute(relative) ) { throw new Error('Resolved path escapes the storage root'); } return candidate; } ``` 4. **Apply containment checks to every filesystem sink** Use the safe resolver for directory creation, L1 through L4 loading and saving, health checks, and synchronization operations. Do not rely on validation at only one call site. 5. **Reject absolute paths and traversal components** Explicitly reject identifiers containing path separators, `.` or `..` path components, null bytes, and platform-specific alternate separators. 6. **Use tenant isolation and least privilege** Run the service under a dedicated account with access only to its memory root. For multi-tenant deployments, use separate roots or operating-system i ...[truncated 341 chars]
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • 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)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding
声明描述的是一个功能完整、可自动运行的商业化多智能体系统,重点在全网监控、内容生产、趋势分析、自动执行和定时报告。实际代码只是“多智能体协作系统演示”,主要用于展示代理注册、任务分发、协作、向人求助和简单对话。它没有外部网络访问、没有真实数据源接入实现、没有内容生成流水线、没有定时器或调度器、没有主动感知机制,也没有反思/自适应逻辑。虽然代码与“多智能体协作”这一高层概念相关,但与声明中的主要业务能力和自动化程度存在明显且实质性的差距,因此构成描述与行为不匹配。

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding
The description markets a broad autonomous business agent team with monitoring, creation, execution, adaptation, and scheduled reporting. The actual code chunk is only a detailed console demo of an underlying library's features, using static sample data. It demonstrates memory management, signal scoring/classification, workflow summarization, and personal goal reflection, which partially align with concepts like trend/signal analysis and automatic record-keeping. However, the main advertised capabilities—web collection, content generation, autonomous task execution, scheduled report triggers, and orchestration features like intent recognition and smart routing—are not present in the code shown. Therefore the declared description materially overstates and differs from the supplied code behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding
The description promises a comprehensive multi-agent system with external monitoring, content production, trend prediction, proactive automation, and scheduled reporting across industries. The supplied code chunk does not implement those end-user capabilities. Instead, it exposes interfaces and methods for a layered memory subsystem: storing variables/context, adding/querying memories, archiving between levels, syncing memory entries between systems, and generating reflective insights from memory. Reflection-related methods like generateMirrorInsight loosely align with the claimed '反思机制', but they are only a small supporting component. The code's primary purpose is internal memory management, which is materially narrower and different from the declared product behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding
The declared description promises a broad autonomous multi-agent business system that monitors industry hotspots across the web, creates content, predicts trends, routes tasks intelligently, adapts to users, executes work automatically, and sends scheduled reports. The actual code chunk does none of that directly. It only provides a persistence and memory layer: local filesystem-backed storage for working/experience/knowledge/wisdom memories, simple search, archival, sync files, summary generation, and health checks. While a memory subsystem could support a larger agent platform, this chunk’s actual behavior is materially narrower and different from the marketed primary purpose. No evidence of web collection, scheduling, reporting, content generation, industry-specific workflows, or autonomous task execution appears in the supplied code.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding
The declared description promises a broad, fully featured multi-agent automation platform with monitoring, analysis, content generation, autonomous execution, and scheduled reports. However, the provided code chunk contains no operative code at all—just an empty declaration file. Therefore the actual behavior does not substantiate any of the claimed capabilities, making the description materially inconsistent with the supplied code.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding
The declared description promises a production-style autonomous multi-agent system with broad business capabilities: internet-wide collection, content generation, trend prediction, automatic logging/execution, adaptive routing, and scheduled reporting. The supplied code chunk is instead a console example file that exercises methods on an imported AICollaborationSystem using static sample data. It demonstrates signal classification, workflow knowledge capture, goal tracking, insight generation, and health reporting at a toy/example level. There is no evidence in this chunk of network access, cross-industry monitoring, automated report scheduling, content generation pipelines, autonomous execution, or the claimed agent orchestration features. This is a material description-behavior mismatch, with the actual code serving as a usage demo rather than implementing the advertised autonomous skill.

Tp4

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding
The declared description promises a broad, autonomous, multi-agent system with web-scale monitoring, content generation, trend prediction, auto-execution, adaptive behavior, and scheduled reports. The supplied code chunk is only a declaration file describing an integrated class that wraps memory, signal recognition, workflow, and goal systems. Its exposed methods suggest internal coordination, querying, daily scans, workflow processing, and goal review, but do not substantiate the specific marketed capabilities in the description. In particular, there are no triggers, no scheduling, no declared external/network access, and no explicit APIs for content creation, web monitoring, report delivery, or autonomous action. This is a material description-to-behavior mismatch, with the implementation appearing much narrower and more generic than claimed.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding
The declared description promises a broad autonomous multi-agent system with industry-wide monitoring, content creation, trend forecasting, proactive automation, and fixed-time scheduled reporting. The provided code chunk is only a top-level integration class that initializes four local modules: memory, signal recognition, workflow assets, and personal goals. Its methods mainly delegate to those modules for summaries, health checks, scans, workflow reports, and weekly self-review. There is no visible implementation of external web collection, scheduled execution, autonomous action-taking, or the named advanced behaviors. Additionally, the code emphasizes personal goal/self-awareness tracking, which is materially different from the marketed '赚钱' hot-topic/content automation framing. Therefore the description overstates and misrepresents the observed behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding
The declared description markets a multi-agent, cross-industry autonomous business intelligence and content production system that monitors the web, predicts trends, executes tasks automatically, and sends scheduled reports. The supplied code chunk instead defines interfaces and methods for a personal goal tracking system focused on motivation analysis, goal networks, energy/time allocation, blind-spot discovery, self-reflection letters, future-self prediction, and weekly self-awareness reporting. This is a materially different primary purpose. The code does not show the advertised agents, web collection, content creation, trend monitoring, automatic execution, or scheduled reporting behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding
The description advertises a broad commercial automation product with four AI agents handling internet-scale data collection, content strategy, trend prediction, and automatic execution across many industries, including scheduled reporting. The supplied code instead defines a PersonalGoalSystem focused on individual self-management: motivation analysis, goal conflict/synergy mapping, energy allocation analysis, cognitive blind-spot discovery, reflective weekly reports, and generic future predictions. It mainly reads/writes from a memory subsystem and syncs internal goal/value data to two other systems. There is no evidence of external web scraping, industry monitoring, content production, autonomous business workflows, or scheduled triggers. This is a strong description-behavior mismatch, with a materially different primary purpose.

Tp4

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding
The description markets a comprehensive autonomous multi-agent business system with web-wide monitoring, content generation, forecasting, automatic execution, adaptive routing, and fixed-time reporting. The supplied code chunk is much narrower: it only exposes a signal recognition/reporting component API. It can classify signals, generate a daily scan report, identify patterns, query stored signals, and sync with other systems through memory, which partially aligns with monitoring/reporting. However, there is no evidence here of content creation, agent coordination, scheduled execution, proactive collection, or broad autonomous task execution. Because the actual code reflects only a subset of the declared platform behavior and lacks several headline capabilities, the description does not accurately represent this code chunk.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding
The declared description presents a comprehensive autonomous multi-agent system with broad business automation capabilities. The supplied code chunk is only a single 'signal recognition' module. Its primary behavior is to evaluate structured signal objects using fixed heuristics, save higher-value signals into memory, generate a report from already-supplied signals, identify simple group patterns, and sync some stored items to other systems. This is materially narrower than the declared purpose. While the report-generation function loosely relates to monitoring/trend insight, the major advertised capabilities—full-network collection, content creation, automatic execution, intent routing, proactive sensing, adaptive behavior, and scheduled reporting—are not implemented in this chunk. Therefore the description does not accurately represent what this code actually does.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding
The description promises a large-scale autonomous multi-agent system with cross-industry monitoring, content production, forecasting, and automatic execution. The actual code chunk is a declaration file for a workflow asset/knowledge management subsystem. Its methods revolve around structuring tacit knowledge, capability extraction, methodology construction, and report generation from provided inputs. There is no evidence in this chunk of network/data collection, content generation, forecasting models, scheduling, trigger handling, or automation across industries. While 'generateDailyWorkflowReport' loosely overlaps with reporting, it is input-driven and does not establish the claimed twice-daily autonomous reporting behavior. Therefore the declared purpose materially overstates and misrepresents the code's actual behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding
The declared description promises a large-scale autonomous AI workforce with four specialized agents, industry-wide monitoring, content generation, trend insight, and automatic execution with scheduled reports. The supplied code does not implement any of those primary behaviors. Instead, it is narrowly focused on internal workflow knowledge extraction and storage: recording user responses at several abstraction levels, deriving capability patterns, constructing a methodology artifact, querying stored methodologies, and syncing them to other systems. There is no evidence of internet collection, monitoring, content generation, prediction, scheduling, trigger handling, autonomous execution, or adaptive routing. This is a material purpose mismatch, not merely an incomplete excerpt of supporting logic, because the implemented subsystem serves a different function than the advertised end-user capability set.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding
The declared description promises a comprehensive autonomous multi-agent system with broad industry applicability and operational automation. The supplied code chunk is much narrower: it is a console demo illustrating a holiday duty mechanism for hotspot monitoring and reporting using static example data and memory storage. While there is some loose thematic overlap around hotspot monitoring and reporting, the main advertised capabilities—real collection across the web, multi-agent collaboration, content creation, forecasting, autonomous execution, adaptive behavior, and timed automatic reporting—are not actually implemented in this code. Therefore the description materially overstates and misrepresents the code's real behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding
The description markets a comprehensive autonomous multi-agent system with broad cross-industry monitoring, creation, prediction, adaptive routing, and timed automatic execution. The supplied code chunk does not implement that overall behavior. It is explicitly labeled as a scenario example, not a standalone system, and mainly demonstrates hotspot scoring, level classification, report text generation, memory storage, and sample calls into an imported collaboration system. There is no evidence in this chunk of web-scale collection, automatic scheduled runs, content generation beyond templated reporting, forecasting, or the advanced agent orchestration features claimed in the description. Therefore the declared description materially overstates and differs from the actual code behavior.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding
The declared description promises a multi-agent autonomous business system spanning monitoring, content generation, forecasting, adaptive execution, and scheduled reports across industries. The actual supplied code chunk does not implement those features; it only performs environment setup for a Node/TypeScript project and initializes local folders. While installation is a supporting detail, this chunk’s primary behavior is materially different from the declared end-user functionality, and none of the headline capabilities or timing behavior are evidenced here.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding
The declared description presents a broad production-style AI workforce with monitoring, creation, prediction, automation, and scheduled reporting across industries. The supplied code instead is a console-based demonstration of a memory-system design: scenario configuration, mock memory records with metadata, simulated usage tracking, scoring, and printed API examples. There are no network calls, no data ingestion from the web, no content generation logic, no trend-analysis engine, no scheduler, no triggers, and no real execution of the advertised multi-agent workflow. This is not merely an incomplete implementation detail; the primary purpose of the code chunk is materially different from the declared skill purpose.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding
The declared description promises a comprehensive autonomous multi-agent business system with broad capabilities across monitoring, creation, prediction, adaptation, and automated reporting. The actual code does not implement those features. Instead, it is a demonstration script focused narrowly on a news time-decay scoring policy: calculating age-based decay, downgrading older items, filtering visibility, sorting by freshness, printing results, and saving a config object to memory. There are no triggers, schedulers, network collection, content generation, trend prediction, routing, reflection, or autonomous execution behaviors shown in this code chunk. This is a material mismatch in primary purpose and implemented capabilities.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding
The description presents a sophisticated autonomous multi-agent skill with broad industry coverage, active monitoring, prediction, content generation, adaptive behavior, and scheduled automatic reporting. The supplied code is much narrower: it is a demo script for evaluating a predefined list of hotspot items, generating a textual report, saving summaries to an internal memory subsystem, and printing analysis scaffolding. While it loosely aligns with 'hotspot monitoring' and report generation, most of the headline capabilities are not implemented in the provided chunk. The primary purpose is therefore materially narrower than declared, and several key claimed behaviors are absent.

Ssd 3

High
Confidence
97% confidence
Finding
Cross-system synchronization serializes collected memory entries into shared files with no authentication, authorization, sensitivity filtering, or encryption. This enables broad redistribution of accumulated user and operational data across components or systems, making any previously stored secret or regulated data much easier to expose at scale.

Description-Behavior Mismatch

High
Confidence
98% confidence
Finding
The manifest describes a multi-agent business automation skill that monitors all-industry热点, creates content, predicts trends, auto-records work, and automatically executes tasks on a schedule. This file instead implements a personal self-reflection and goal-tracking system: motivation analysis, energy allocation, blind-spot detection, weekly self-awareness reports, and syncing personal goal data to other internal systems.

Missing User Warnings

High
Confidence
98% confidence
Finding
The reset instructions recommend deleting the memory directory with no explicit warning that the action is irreversible and may destroy all accumulated records, logs, and user data. Users may execute it verbatim and suffer permanent data loss, especially because the skill heavily emphasizes automated persistence and long-term memory storage.

Tool Parameter Abuse

High
Category
Tool Misuse
Content
**A**:
```bash
# 删除记忆目录
rm -rf memory/my_system/

# 重新创建系统
const ai = new AICollaborationSystem('my_system');
Confidence
97% confidence
Finding
The documentation includes a destructive shell command, `rm -rf memory/my_system/`, that can erase data recursively if copied and run by users. In agent or terminal-assisted contexts, such commands are especially dangerous because users may execute them without understanding path scope, and small modifications or path confusion could broaden deletion beyond the intended directory.

Natural-Language Policy Violations

Medium
Confidence
94% confidence
Finding
The entire skill prompt file is written as Chinese-only operating instructions and output templates, with no indication that users may choose another language or locale. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation.

Static analysis

No suspicious patterns detected.