The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides detailed instructions and code for accessing sensitive browser profile directories (e.g., ~/Library/Application Support/Google/Chrome or %LOCALAPPDATA%\Google\Chrome) to export authentication cookies and transfer them to cloud environments.
[DYNAMIC_EXECUTION]: The skill implements a pipeline (SCRIPT-CODEGEN.md) that uses LLMs to generate Playwright or Stagehand scripts from task traces, which are then executed. Additionally, the browse CLI tool supports an eval command for executing arbitrary JavaScript within the browser context.
[COMMAND_EXECUTION]: The skill describes patterns for spawning detached background processes (spawn with {detached: true}) to monitor CDP event streams and manage browser session tracing, involving PID management and signal handling.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from arbitrary, untrusted websites as accessibility trees, HTML, and markdown. While it includes defensive measures like domain allowlists, it maintains high-privilege capabilities to interact with these pages (clicking, typing, form submission).
Ingestion points: Accessibility tree snapshots, HTML body extraction, and markdown conversion of remote web pages in SNAPSHOT-AND-ACT.md and SAFE-BROWSER.md.
Boundary markers: Implements a domain allowlist firewall and explicitly labels page content as untrusted in SAFE-BROWSER.md.
Capability inventory: Full browser interaction (click, fill, type, upload), script generation and execution, and eval capabilities across all automation scripts.
Sanitization: Uses a CDP-based firewall to intercept and fail off-domain network requests at the browser level in SAFE-BROWSER.md.
[EXTERNAL_DOWNLOADS]: The skill references and downloads scripts, configuration, and tools from external sources, including the browserbase and agent-browser ecosystems. These references target well-known repositories and package registries.

browser-automation