seo-image-gen

• Outcome: Review
• Latest audit: Aug 12, 2026 · 2:56 PM UTC
• Version: 1.0.0

Overview

This SEO image-generation skill is mostly coherent, but it needs review because it gives safety-filter rephrasing guidance and broad direct API/external-service instructions that are under-scoped for a public skill.

Review this skill before installing. It is not destructive and contains no executable code in the artifact, but it relies on external Gemini/nanobanana services, may store generated outputs and cost data locally, and includes safety-filter workaround guidance that should be removed or rewritten to require safe redirection rather than bypass-style rephrasing.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (9)

Context-Inappropriate Capability

Medium

• Confidence: 96% confidence
• Finding: The documentation explicitly advises users to bypass the MCP skill and call the underlying Google API directly with curl or the SDK if a feature is unsupported. That undermines any control, auditing, guardrails, or policy enforcement provided by the skill/MCP layer and can lead agents or users to operate outside the intended security boundary.

Vague Triggers

Medium

• Confidence: 90% confidence
• Finding: The skill advertises broad, common-language triggers such as "generate image", "banner", and "thumbnail", which can cause the skill to activate in situations where the user did not clearly intend to send content to an external image-generation service. In this skill’s context, unintended invocation matters because prompts, brand details, and possibly sensitive page/content descriptions may be transmitted to external MCP-backed tools without an explicit privacy warning at the trigger point.

Missing User Warnings

Medium

• Confidence: 96% confidence
• Finding: The skill description explains capabilities but does not clearly warn that user prompts and related content are sent to external image-generation services via Gemini/nanobanana MCP. This creates a data-exposure risk because users may provide unpublished marketing copy, product concepts, campaign plans, or brand assets under the assumption processing is local or first-party.

Missing User Warnings

Medium

• Confidence: 89% confidence
• Finding: The document states that a ledger is stored at '~/.banana/costs.json' and created automatically on first use, but it does not mention this storage behavior to the user or require consent. Silent local persistence can expose usage history, prompts, filenames, or cost metadata to other local users, backups, or later processes, creating a privacy and transparency risk.

Missing User Warnings

Low

• Confidence: 82% confidence
• Finding: The tool documentation states that generated images are automatically saved to a fixed directory under the user's home folder, but it does not warn about persistent local file creation. This can expose sensitive or regulated visual content to other local users, backups, indexing tools, or later unintended reuse because users may assume outputs are ephemeral.

Missing User Warnings

Medium

• Confidence: 94% confidence
• Finding: The documentation requires a Google AI API key and describes cloud-backed image generation features, but it does not disclose that prompts and image content will be transmitted to an external Google service. This omission can cause users to send confidential, personal, or proprietary data off-device without informed consent or organizational approval.

Missing User Warnings

Medium

• Confidence: 95% confidence
• Finding: This section explicitly teaches users how to rephrase prompts after safety blocks without any warning that blocked outputs may reflect policy or safety constraints. In a prompt-engineering reference for an image-generation skill, that omission normalizes bypass behavior and can encourage users or downstream agents to preserve unsafe intent while iterating around safeguards.

Ssd 1

High

• Confidence: 98% confidence
• Finding: The 'Safety Filter Rephrase Strategies' section instructs how to keep moving toward blocked content by changing wording, framing, or context instead of abandoning unsafe requests. That is a classic moderation-evasion pattern: it preserves disallowed intent while helping an operator probe for phrasings that slip through filters, which can enable generation of harmful or policy-violating imagery.

Ssd 2

High

• Confidence: 97% confidence
• Finding: The examples provide concrete paraphrases for converting blocked prompts into variants more likely to pass moderation, including violence, medical/gore, public figures, minors, and suggestive content. These are actionable evasion examples that can be copied directly, lowering the barrier to bypassing safety systems and making misuse more practical.

VirusTotal

VirusTotal findings are pending for this skill version.

Static analysis

No suspicious patterns detected.