The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes specific configuration instructions, such as a preference for using "gemini-flash-latest" and avoiding "gemini-1.5-flash". These are functional constraints intended to guide model selection and do not constitute a malicious attempt to override safety filters or extract system prompts.
[EXTERNAL_DOWNLOADS]: The documentation references the installation of the official Firebase CLI (npm install -g firebase@latest) and provides links to official Firebase documentation domains (firebase.google.com). These references target well-known, trusted resources associated with the skill's stated vendor.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates how to process untrusted data from users (prompts) and external files (images, audio, video, PDFs) through generative models.
Ingestion points: User-provided prompt strings in text generation functions and imageFile objects in multimodal functions within references/usage_patterns_web.md.
Boundary markers: The provided code snippets do not include explicit delimiter-based boundary markers or instructions to ignore embedded content.
Capability inventory: The skill utilizes generateContent, sendMessage, and generateContentStream to process these inputs.
Sanitization: The examples show raw data passing to the model without explicit sanitization or validation steps.
Risk Assessment: While this creates an attack surface for indirect prompt injection, it is inherent to the skill's purpose of providing AI integration examples and is documented here for situational awareness.

firebase-ai-logic