Back to skill

Security audit

艺术

Security checks for vulnerabilities and agentic risk

Overview

This skill is framed as art guidance but asks for broad file and command powers that are not clearly scoped, so it should be reviewed before installation.

Treat this as an automation-capable skill, not just an art-advice prompt. Install it only in an environment where exec and write are disabled or require explicit per-action approval, and restrict any file reads or writes to user-selected artwork or project files.

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:35
Finding

Excessive Tool Permissions for an Art-Guidance Skill

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:35-38
Vulnerability Type: Excessive tool permissions and violation of least privilege
Risk Level: Medium

yaml
tools:
  - read
  - exec
  - write

Technical Analysis

The skill declares read, write, and exec capabilities despite its stated purpose being to provide art-creation guidance, technique development, and appreciation advice. General command execution and unrestricted filesystem modification are not demonstrably required for that purpose.

The document does not define a command allowlist, filesystem path restrictions, user-confirmation requirements, permitted external endpoints, or restrictions on the data that may be processed. Although the document recommends sandboxing and allowlisted commands elsewhere, those recommendations are descriptive and are not represented by enforceable controls in the project.

The audited project contains only SKILL.md; no executable implementation was present. Consequently, no active malicious payload or demonstrated command-injection implementation was found. The risk arises when an Agent platform grants the declared capabilities and interprets the skill's broad processing instructions as authorization to use them.

Attack Path

  1. A user installs or loads the skill into an Agent platform.
  2. The platform grants the skill its declared read, write, and exec tools.
  3. The skill processes user-controlled art content or file-related input.
  4. Without enforceable path, command, or confirmation restrictions, manipulated input may induce the Agent to read unrelated files, modify accessible files, or execute commands outside the legitimate art-guidance task.
  5. Any resulting access is limited by the operating-system privileges, sandbox controls, and credentials of the hosting Agent process.

Impact Assessment

If the declared tools are granted without external sandbox enforcement, the skill ...[truncated 591 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove exec because command execution is not necessary for conversational art guidance.
  2. Remove write unless a documented feature specifically requires creating user-requested output files.
  3. Retain read only when analysis of a user-selected artwork or document is required.
  4. Restrict file access to paths explicitly selected by the user and deny traversal outside the designated workspace.
  5. If writing is retained, require explicit confirmation for each destination and prevent overwriting existing files by default.
  6. If command execution is indispensable, use an enforceable command-and-argument allowlist, prohibit shell interpretation, reject user-controlled command fragments, run commands in an isolated sandbox, and apply operating-system resource limits.
  7. Require per-operation user approval for sensitive tool calls and display the exact path or command before execution.
  8. Restrict network access to explicitly documented HTTPS endpoints and prevent transmission of local files, environment variables, API keys, or other secrets.
  9. Update the capability declaration and documentation so that every requested permission maps to a concrete, narrowly scoped feature.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (5)

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The claim that the skill has removed risky code and external dependencies conflicts with later documentation advertising API integration and command execution. This misleading safety language is dangerous because it can reduce reviewer and user scrutiny, causing over-trust in a skill that still requests or describes powerful capabilities.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill is presented as an art-guidance assistant, but the documentation exposes broad automation capabilities including file access, API use, and command execution. This mismatch increases the risk that users or host agents grant powerful privileges under a benign pretense, enabling abuse of local resources or external systems if the skill is invoked in permissive environments.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Documenting system command execution for an art-advice skill is dangerous because exec is a high-risk primitive unrelated to the stated task. In an agent environment, this can be used to run arbitrary shell commands, access sensitive local data, pivot to other tools, or alter the host system, especially if user input is ever incorporated into commands.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

File read/write support is broader than necessary for a skill advertised as providing artistic guidance and appreciation. Unnecessary file capabilities increase the chance of unauthorized reading of local files, overwriting user content, or being used as part of a larger exfiltration or persistence workflow inside an agent runtime.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

External API integration is not clearly justified by the art-focused purpose and expands the attack surface by allowing outbound data flow. In agent settings, undocumented or weakly justified network access can facilitate prompt exfiltration, leakage of user content, or interaction with untrusted third-party services.

Content

No source excerpt is available for this finding.