Back to skill

Security audit

Deepseek聊天

Security checks for vulnerabilities and agentic risk

Overview

This DeepSeek chat skill is not clearly malicious, but it asks for broad read, write, and command-execution authority that does not fit its stated chat/API purpose.

Review this skill before installing. Use it only in a tightly sandboxed agent environment, avoid granting write or command execution unless you have a specific need, and do not send sensitive files or credentials through it unless the data flow and DeepSeek API use are clear to you.

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:25
Finding

Excessive File-System and Command-Execution Permissions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 25–28; related capability claims at lines 240–244
Vulnerability Type: Violation of least privilege through unnecessary Agent tool permissions
Risk Level: Medium

yaml
tools:
- read
- exec
- write

Technical Analysis

The Skill declares access to file-reading, file-writing, and command-execution tools even though its stated primary purpose is to provide DeepSeek chat integration. These system-level capabilities are not necessary for a narrowly implemented chat API client.

The related capability overview at lines 240–244 explicitly advertises arbitrary file processing, external API integration, and system-command execution. However, the package contains no implementation that restricts accessible paths, defines an allowed-command list, validates command arguments, requires confirmation before side effects, or enforces the claimed sandbox.

Tool declarations do not independently prove that a host Agent will grant unrestricted access. Nevertheless, requesting all three capabilities materially expands the Skill's potential authority and violates least-privilege design. If the hosting Agent trusts these declarations or follows subsequent requests under the Skill's context, the exposed tools could be used for operations unrelated to chat.

Attack Path

  1. A user installs or loads the Skill in an Agent platform that honors its requested tools.
  2. The platform grants read, write, and exec capabilities based on the Skill manifest.
  3. A request processed in the Skill's context causes the Agent to read an arbitrary local file, write to an unintended path, or execute a system command.
  4. Because the Skill defines no path constraints, command allowlist, argument validation, or mandatory confirmation boundary, the Agent may perform the operation with its own operating-system privileges.
  5. Local information could consequently be disclosed or system ...[truncated 1138 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove read, write, and exec from the tool declaration unless each capability is essential to an implemented and documented workflow.
  2. For a chat-only integration, expose a narrowly scoped HTTPS client restricted to the required official DeepSeek API origin and approved request methods.
  3. If file access is genuinely required, restrict it to explicitly approved project-relative paths, reject traversal and symbolic-link escapes, and separate read-only from write access.
  4. If command execution is unavoidable, replace general-purpose exec access with fixed operations and an explicit executable-and-argument allowlist. Do not invoke a shell with user-controlled text.
  5. Require clear user confirmation before any file modification, external transmission of local data, or command with side effects.
  6. Enforce sandbox restrictions in executable policy or host configuration rather than merely claiming a sandbox in documentation.
  7. Add an actual implementation and security tests proving that disallowed paths, commands, endpoints, and arguments are rejected.
  8. Align the documented capabilities with the implemented functionality and remove unsupported claims about file processing and command execution.
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 (4)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The activation guidance is overly broad, using generic triggers like improving efficiency, automation, batch processing, and workflow optimization. This can cause the skill to be selected for many unrelated tasks, which is dangerous here because the skill also advertises external API use and potentially impactful operations, increasing the chance of unintended data exposure or execution.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill is presented as a DeepSeek chat/API tool, but its documentation expands into unrelated capabilities such as bulk messaging, variable injection, archiving, file handling, and command execution. This mismatch can mislead users and downstream agents about the skill’s true behavior, increasing the risk of inappropriate invocation, over-privileged use, or unsafe handling of data and system actions.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The FAQ and feature sections claim support for media messaging, batch sending, storage, and other platform-like capabilities that do not fit a simple chat API skill. Such fabricated or inflated capability claims create a security problem because an agent may route sensitive content, assume persistence or transport controls exist, or invoke the skill for tasks it is not designed to perform.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill advertises file writing, command execution, and external API interaction without prominent upfront warnings about system modification, data transmission, or permission impact. In agent environments, this is especially risky because vague descriptions combined with powerful tools can lead to silent execution, unintended file changes, or sending sensitive content to external services.

Content

No source excerpt is available for this finding.