The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill exposes an execute function designed to run arbitrary JavaScript code on the targeted web pages. While this is a core feature for web automation and data extraction, it allows the agent to execute unverified logic in a remote browser environment.
[EXTERNAL_DOWNLOADS]: The documentation instructs users to install the belt-sh/cli tool using npx and references a setup script hosted at raw.githubusercontent.com/inference-sh/skills/. These are required components for the skill's functionality.
[COMMAND_EXECUTION]: The skill uses the Bash tool to run the belt CLI, which serves as the interface for all browser interactions. Documentation also suggests using oathtool for TOTP generation in authentication scripts (references/authentication.md).
[DATA_EXFILTRATION]: The skill includes capabilities to extract browser cookies, session data, and page content (text, screenshots, video). These features are intended for web scraping and session persistence but allow for the retrieval of sensitive information from the browser context.
[PROMPT_INJECTION]: The skill represents an indirect prompt injection surface as it ingests untrusted data from web pages.
Ingestion points: Page titles, body text, and element descriptions are retrieved via the open, snapshot, and execute functions (SKILL.md, references/commands.md).
Boundary markers: The skill does not implement delimiters or safety warnings for the ingested content.
Capability inventory: The skill possesses capabilities for element interaction (click, fill), JavaScript execution, and file uploads across all automation scripts (references/commands.md).
Sanitization: There is no evidence of filtering or sanitization of external content before it is returned to the agent context.

agent-browser