The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses child_process.spawn in scripts/gemini-webapi/utils/load-browser-cookies.ts to launch browser executables (Chrome, Edge, or Chromium) found on the host system to facilitate authentication.
[CREDENTIALS_UNSAFE]: The skill extracts sensitive session cookies including __Secure-1PSID and __Secure-1PSIDTS from the automated browser and saves them in plain text JSON files within the user's application data directory (e.g., ~/.local/share/baoyu-skills/gemini-web/cookies.json).
[DATA_EXFILTRATION]: The skill reads local files provided via command-line arguments and uploads their content to Google's content-push.googleapis.com endpoint during image generation and vision tasks.
[EXTERNAL_DOWNLOADS]: The skill downloads generated or web-sourced images from Google's googleusercontent.com domains directly to the local filesystem using fetch in scripts/gemini-webapi/types/image.ts.
[PROMPT_INJECTION]: The skill incorporates external data from local files into AI generation prompts, creating a surface for indirect prompt injection.
Ingestion points: Prompt content and reference images are read from the filesystem in scripts/main.ts via readFile.
Boundary markers: No specific delimiters or safety instructions are used when concatenating file content into the API request body in scripts/gemini-webapi/client.ts.
Capability inventory: The skill has permissions for network communication (Google APIs), broad file system access (reading prompts/images, writing cookies/sessions), and process execution (launching browsers).
Sanitization: There is no evidence of sanitization, filtering, or validation of the content retrieved from external files before it is processed by the AI model.

baoyu-danger-gemini-web