The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill extracts highly sensitive Google session cookies, including __Secure-1PSID and __Secure-1PSIDTS, from local browser profiles (Chrome, Chromium, Edge). These credentials provide full access to the user's Google account context for the Gemini service. The extracted cookies are stored locally in a JSON file (cookies.json) in plain text.
[COMMAND_EXECUTION]: To retrieve cookies, the skill launches browser instances with remote debugging enabled (--remote-debugging-port) or connects to already running instances. Additionally, it executes shell commands via execSync to resolve file paths in environments like WSL (e.g., calling cmd.exe and wslpath).
[METADATA_POISONING]: The SKILL.md file describes a mandatory 'Consent Check' flow involving a consent.json file and a user disclaimer. However, the provided script implementation (main.ts) skips this process entirely and proceeds directly to cookie extraction and API initialization. This discrepancy between documented safety protocols and actual behavior is a significant security concern.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from local files provided via the --promptfiles and --reference flags. These inputs are sent to the AI model without visible sanitization, creating a surface for potential prompt injection attacks if the ingested files contain malicious instructions.
Ingestion points: Files read in scripts/main.ts and vision inputs processed in scripts/gemini-webapi/client.ts.
Capability inventory: The skill possesses capabilities for network communication (Google APIs), file system writing (saving images and sessions), and local command execution.
Boundary markers: None observed in the prompt construction logic.
Sanitization: Content from external files is concatenated and passed directly as the model prompt.
[EXTERNAL_DOWNLOADS]: The skill utilizes the baoyu-chrome-cdp package for browser interaction. This dependency is authored by the same vendor as the skill itself. The skill also makes numerous network requests to official Google domains (gemini.google.com, accounts.google.com) to function.

baoyu-danger-gemini-web