Back to skill

Security audit

图谱

Security checks for vulnerabilities and agentic risk

Overview

This skill appears to manage local knowledge notes, but it asks for broad command execution and API-style behavior without clear limits.

Review this carefully before installing. It does not show malware or persistence, but it should be treated as higher risk because it can request command execution while only documenting vague sandbox protections. Install only in a constrained workspace, avoid exposing sensitive notes or API keys, and prefer a version that removes exec or documents exact allowed commands and confirmation rules.

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

Unnecessary Command-Execution Permission Violates Least Privilege

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:21-24
Vulnerability Type: Excessive system-tool permission
Risk Level: Medium

Vulnerable Code

yaml
tools:
- read
- exec
- write

The command-execution capability is also advertised at SKILL.md:198, although the project provides no implementation that enforces the claimed sandbox.

Technical Analysis

The Skill declares the unrestricted exec tool even though its documented purpose is to add, remove, replace, and organize knowledge notes. That workflow requires controlled file reading and writing but does not establish a legitimate need for arbitrary system-command execution.

The package consists only of SKILL.md. It contains no executable implementation, command allowlist, argument validation, shell-isolation policy, or sandbox configuration. Consequently, the documentation's claim that commands run in a secure sandbox is not enforced by anything in the audited project.

If the hosting Agent grants tools according to this declaration, the Skill receives a capability beyond its stated functional requirements. Untrusted note content or task instructions could then influence the Agent into invoking operating-system commands.

Attack Path

  1. A user or automated workflow loads the Skill.
  2. The hosting Agent grants the declared read, write, and exec tools.
  3. The Skill processes attacker-controlled note content, metadata, or task instructions.
  4. Malicious content directs the Agent to perform a system operation as part of the purported note-maintenance workflow.
  5. Because this package defines no enforceable command allowlist or argument restrictions, the Agent may invoke exec with attacker-influenced commands.
  6. The command runs with the operating-system privileges and environmental access available to the hosting Agent.

Impact Assessment

Successful exploitation could permit command execution within the Agent's run ...[truncated 618 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove exec from the declared tools because it is not required for the stated knowledge-note workflow:
    yaml
    tools:
    - read
    - write
    
  2. Restrict read and write operations to the documented life/areas/** directory and reject path traversal, absolute paths, and symbolic-link escapes.
  3. If command execution is genuinely necessary, document each required command and expose a narrowly scoped operation rather than unrestricted shell access.
  4. Implement a strict command allowlist using fixed executable paths and structured argument arrays. Do not concatenate user-controlled values into shell command strings.
  5. Disable shell interpretation and prohibit command substitutions, pipes, redirections, metacharacters, and environment-variable expansion.
  6. Enforce sandboxing at runtime with minimal filesystem permissions, no unnecessary network access, a sanitized environment, resource limits, and a non-privileged operating-system account.
  7. Require explicit user confirmation before any command with side effects and record security-relevant executions in an audit log.
  8. Add automated tests confirming that unapproved commands, path escapes, and attacker-controlled arguments are rejected.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • 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
Findings (5)

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
72% confidence
Finding

The core description frames the skill as maintaining notes/knowledge graph content under a local path, which implies local file operations. Elsewhere the file states that an API key is required and that the skill integrates with external services, creating a contradiction in the documented intent about whether this is a local graph-maintenance skill or a network-dependent external-service skill.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The description says to use the skill whenever AI model calling, intelligent dialogue, agent orchestration, or LLM applications are needed. This scope is extremely broad and does not clearly distinguish when this specific knowledge-graph skill should activate versus when other general-purpose skills should be used.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill exposes read, write, and exec capabilities near the top-level description without an upfront warning that it can modify local data or run commands. In agent ecosystems, insufficiently prominent disclosure can lead users or orchestrators to invoke the skill without informed consent, increasing the chance of unintended file changes or command execution in sensitive repositories.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The FAQ says the skill currently supports Chinese interaction, which imposes a language restriction in natural language. The file does not frame this as an opt-in preference or explain a necessary region-specific reason for the restriction, so it appears to force a specific language policy on users.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill advertises command execution as occurring in a safe sandbox with whitelist controls, but the file exposes a generic exec capability without any enforceable constraints or concrete whitelist definition. This mismatch can mislead users and downstream agents into granting excessive trust, enabling arbitrary command execution or unsafe side effects under the guise of a restricted environment.

Content

No source excerpt is available for this finding.