Back to skill

Security audit

深度研究引擎

Security checks for vulnerabilities and agentic risk

Overview

This research skill is not clearly malicious, but it asks for command execution and gives privilege-related guidance that is broader than its stated research purpose needs.

Install only if you are comfortable with an execution-capable research skill. Run it in a low-privilege sandbox, avoid administrator/root sessions, do not expose sensitive API keys or private files unless required, and require explicit approval before any command execution, file writing, callback/network use, or permission change.

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:21
Finding

Excessive Command-Execution Capability and Privilege-Elevation Guidance

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:21-25; related privilege-elevation guidance at SKILL.md:355
Vulnerability Type: Excessive tool permissions and violation of least privilege
Risk Level: Medium

Vulnerable Code

yaml
tools:
- read
- exec
- glob
- grep

The skill also advises users experiencing insufficient permissions to check file permissions and run with administrator privileges at SKILL.md:355.

Technical Analysis

The declared research workflow primarily requires information retrieval, content reading, source comparison, and report generation. Arbitrary system-command execution is not demonstrated as necessary for those functions. Nevertheless, the skill requests access to the unrestricted exec capability and separately recommends administrator execution when permission errors occur.

This configuration violates the principle of least privilege. If user-controlled input, malicious research content, or indirect prompt injection influences commands generated by the agent, the available exec capability may turn untrusted natural-language instructions into local system operations. Running the agent with administrator privileges would increase the resulting commands' authority.

The reviewed file does not contain a concrete malicious command or an implementation that automatically elevates privileges. Exploitation therefore depends on the hosting agent permitting command execution and allowing untrusted content to influence tool calls.

Attack Path

  1. An attacker supplies a crafted research request or causes the agent to retrieve a malicious source containing instructions designed to influence agent behavior.
  2. The agent processes the untrusted content while the skill has access to the exec tool.
  3. The malicious content induces the agent to construct and execute an unintended system command.
  4. If a permission error occurs, the skill's troubleshooting guidance ma ...[truncated 828 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove exec from the declared tools unless a documented, essential research operation cannot be completed through safer capabilities.
  2. Restrict the skill to read-only retrieval and analysis tools such as read, glob, and grep.
  3. Remove the recommendation to run the agent with administrator privileges. Troubleshooting guidance should instead direct users to grant only the minimum file or resource permission required.
  4. If command execution is essential, replace unrestricted execution with a narrowly defined allowlist of fixed commands and permitted arguments.
  5. Reject shell metacharacters and prevent user-controlled values from being interpolated into command strings.
  6. Execute commands in a sandbox with restricted filesystem access, no inherited secrets, limited network connectivity, resource limits, and an unprivileged operating-system account.
  7. Treat retrieved source content as untrusted data and explicitly prohibit instructions embedded in research sources from authorizing tool calls.
  8. Require user confirmation before any command that writes files, accesses sensitive locations, invokes the network, or changes system state.
  9. Record command invocations and security decisions in an audit log without exposing credentials or other sensitive values.
Vulnerability Patterns
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The description explicitly states "支持中文交互" as a built-in behavior, and the manifest display name/content consistently present the skill as Chinese-only. There is no indication that users can choose another language or opt in to this locale constraint, which makes it a natural-language policy concern.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill metadata and body give conflicting statements about whether execution is supported, including 'MD+EXEC' in one section while later describing the skill as pure Markdown-driven. In an agent environment, capability ambiguity is security-relevant because operators may approve or invoke the skill under the wrong trust assumptions, potentially enabling command execution when the user expected a non-executing research-only skill.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The document claims file handling support including write operations, but elsewhere its declared capabilities do not consistently support that behavior. This mismatch can mislead users and orchestrators into granting broader filesystem trust than intended, increasing the chance of unsafe automation, accidental overwrites, or policy bypass through incorrect capability assumptions.

Content

No source excerpt is available for this finding.