Back to skill

Security audit

聊天

Security checks for vulnerabilities and agentic risk

Overview

This skill appears to be a chat/API helper, but it asks for broad file and command powers without enough scoping or a working implementation to justify them.

Review this skill carefully before installing. It is not evidence of malware, but it requests powerful local capabilities for a chat-oriented use case and gives unclear API-key instructions. Only use it in a restricted environment, avoid exposing sensitive files or credentials, and require explicit confirmation before any file write or command execution.

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

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:26
Finding

Excessive File-System and Command-Execution Permissions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:26-29
Vulnerability Type: Excessive tool permissions and violation of least privilege
Risk Level: Medium

Vulnerable Code

yaml
tools:
- read
- exec
- write

Related capability claims at SKILL.md:186-190:

markdown
## 主要功能
- **自动化执行**: 豆包大模型对话(免费
- **文件处理**: 支持多种文件格式的读取、解析和写入操作
- **API集成**: 通过标准化接口调用外部服务并处理响应
- **命令执行**: 在安全沙箱中执行系统命令并收集结果
- **信息检索**: 快速搜索和过滤目标数据

Technical Analysis

The skill is presented primarily as a Doubao model conversation and search integration, but it requests general-purpose read, write, and exec tools. Ordinary remote model interaction does not inherently require unrestricted local file access or system command execution.

No implementation, enforceable sandbox policy, command allowlist, file-path restriction, input validation mechanism, or confirmation boundary is included in the audited package. The statement that commands run in a secure sandbox is therefore a documentation claim rather than an enforceable control visible in the project.

If the hosting agent grants these tools directly, attacker-controlled conversation content could influence tool selection and arguments. This creates a confused-deputy risk in which a chat-oriented skill receives capabilities beyond those reasonably necessary for its stated function.

Attack Path

  1. A user or external content supplies a crafted chat instruction to the skill.
  2. The agent loads the skill and grants its declared read, write, and exec capabilities.
  3. The crafted content persuades or causes the agent to invoke one of those tools as part of the requested conversation or automation workflow.
  4. The agent reads a local file, modifies a writable file, or executes a system command using its own runtime privileges.
  5. The resulting data or command output may be returned to the requester, while file changes or command side effects occur within the agent's accessible environment.

The pack ...[truncated 941 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove read, write, and exec unless each capability is strictly necessary for a documented use case.
  2. Replace general-purpose tools with a narrowly scoped HTTPS API client restricted to approved Doubao endpoints, methods, and request schemas.
  3. If local file access is required:
    • Restrict access to a dedicated working directory.
    • Deny traversal, symbolic-link escapes, and absolute paths.
    • Use separate read-only and narrowly scoped write permissions.
    • Require explicit user confirmation before accessing sensitive files.
  4. If command execution is essential:
    • Use a hardened sandbox with no host file-system or credential access.
    • Permit fixed executable paths and structured arguments only.
    • Prohibit shell interpretation and concatenation of user-controlled input.
    • Apply resource, network, environment-variable, and execution-time limits.
    • Require explicit confirmation for commands with side effects.
  5. Treat conversation content, retrieved text, and model output as untrusted data rather than executable instructions.
  6. Document the exact reason each capability is required and test that the skill fails closed when a request exceeds its authorization boundary.
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 (3)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill uses broad 'use when' language covering automation, workflow optimization, API integration, and agent orchestration without tight scope boundaries. Overbroad invocation guidance can cause the agent to select a powerful skill in contexts where file writes, command execution, or external calls are unnecessary, expanding exposure and increasing the chance of unintended side effects.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The documentation is internally inconsistent: it says the LLM API is provided by the agent platform, but also states that a separate API key is required. This can mislead users into supplying credentials unnecessarily or to the wrong execution path, increasing the risk of secret mishandling and unsafe deployment assumptions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill advertises file handling, command execution, and external API integration, but the documentation does not present a clear upfront warning that these actions can affect the local system, modify data, or transmit information externally. In an agent setting, hidden or late-disclosed side effects are dangerous because users may invoke the skill expecting a simple chat function while granting it materially broader capabilities.

Content

No source excerpt is available for this finding.