GPT Image 2 精准文字图片

• Outcome: Pass
• Latest audit: Aug 17, 2026 · 9:43 AM UTC
• Version: 1.0.0

Overview

This skill is an image-generation wrapper that uses a third-party AI Hive API and stores an API key locally, with some leftover generic client code that is broader than the advertised image workflow but not exposed by the active skill CLI.

Install only if you are comfortable using AI Hive as the remote provider, storing an AI Hive API key on disk, and sending prompts plus any reference images to that service. Treat the exact-text claim as requiring manual review, especially for legal, medical, financial, pricing, or certification text.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• MCP Least Privilege: Underdeclared Capability, Wildcard Permission, Missing Permission Declaration
• MCP Tool Poisoning: Hidden Instructions, Unicode Deception, Parameter Description Injection
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration

Findings (8)

Lp3

Medium

• Category: MCP Least Privilege
• Confidence: 83% confidence
• Finding: The skill invokes Python scripts that can use environment variables, file access, shell execution, and network operations, yet the skill declares no permissions. This creates a transparency and governance gap: users or host systems cannot accurately assess what the skill can access, and an overprivileged implementation could exfiltrate data, modify files, or make unauthorized outbound requests without clear disclosure.

Tp4

High

• Category: MCP Tool Poisoning
• Confidence: 96% confidence
• Finding: The documented purpose is narrowly scoped to precise image-text generation and verification, but the underlying implementation reportedly includes unrelated capabilities such as text chat, video generation, model discovery, account/balance access, generic uploads, browser-based API key setup, and generic task retrieval. This mismatch is dangerous because it broadens the attack surface and can mislead users into granting trust to a tool that may access account data, upload arbitrary media, or perform actions outside the advertised scope; additionally, the claimed character-by-character verification appears unsupported, which can cause unsafe reliance in legal, medical, financial, or pricing contexts.

Description-Behavior Mismatch

Medium

• Confidence: 95% confidence
• Finding: The file is marketed as a narrowly scoped GPT Image 2 exact-text image skill, but it actually embeds a broad AI Hive client capable of chat, video generation, model enumeration, account queries, uploads, and task operations. This scope mismatch is dangerous because agents or reviewers may grant it permissions appropriate for an image-only skill while the code can perform materially broader remote actions.

Intent-Code Divergence

Medium

• Confidence: 91% confidence
• Finding: The module docstring explicitly describes a general-purpose AI capability wrapper, which contradicts the claimed fixed-purpose GPT Image 2 skill. This kind of deceptive or misleading packaging increases the chance that downstream systems trust and invoke code under false assumptions about what it can access or do.

Context-Inappropriate Capability

Low

• Confidence: 84% confidence
• Finding: Exposing user-info and wallet/balance retrieval is outside the stated image-generation purpose and creates unnecessary access to potentially sensitive account metadata. In a skill context, extra account-query capability broadens the blast radius if the skill is invoked by an automated agent with API credentials.

Context-Inappropriate Capability

Medium

• Confidence: 94% confidence
• Finding: The skill includes a generic text chat operation unrelated to the advertised exact-text image workflow. That hidden capability allows arbitrary prompt content to be sent to a remote model endpoint, expanding data exfiltration and misuse opportunities beyond what users would reasonably expect from an image-only skill.

Context-Inappropriate Capability

Medium

• Confidence: 94% confidence
• Finding: Video generation support is not justified by the stated image-only purpose and materially expands the operations the skill can perform with uploaded media and remote generation APIs. In agent environments, this hidden breadth can enable unexpected costs, data transfer, and processing of user media under a narrower trust label.

Context-Inappropriate Capability

Low

• Confidence: 78% confidence
• Finding: Generic model-listing is outside the narrow business purpose of this skill and exposes unnecessary discovery capability about the backend platform. While not directly severe, it increases reconnaissance value and broadens what an agent can do with the provided API key beyond exact-text image generation.

VirusTotal

65/65 vendors flagged this skill as clean.

Static analysis

No suspicious patterns detected.