The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill has a surface for indirect prompt injection as it processes external image files and content from URLs.
Ingestion points: scripts/analyze-image.py accepts local file paths and remote URLs for processing.
Boundary markers: The provided documentation does not specify if the agent uses delimiters or explicit 'ignore embedded instructions' prompts when passing data to the API.
Capability inventory: The skill uses Python scripts to perform filesystem reads and network requests to generativelanguage.googleapis.com.
Sanitization: The documentation mentions basic format validation (MIME types, size limits) but does not detail sanitization of the content itself.
[Data Exposure & Exfiltration] (SAFE): The skill implements a standard 3-step lookup for GEMINI_API_KEY. While it searches for .env files in multiple directories (including .claude/), this is a common configuration pattern. The documentation appropriately warns users to add these files to .gitignore to prevent accidental credential leakage.
[External Downloads] (SAFE): The installation instructions recommend pip install google-genai. Because google-genai is the official SDK provided by a trusted organization (Google), this finding is downgraded to SAFE/INFO per the [TRUST-SCOPE-RULE].
[No Code Provided] (SAFE): The primary logic for this skill is described as residing in scripts within the scripts/ directory (analyze-image.py, upload-file.py, etc.). However, these script files were not provided for analysis. The evaluation is based on the provided markdown and configuration templates which are benign.

gemini-vision