The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Documentation in references/quickstart.md suggests executing a remote shell script via curl -fsSL https://browser-use.com/profile.sh | sh to synchronize browser profiles with cloud infrastructure, which constitutes remote code execution from a non-whitelisted domain.
[EXTERNAL_DOWNLOADS]: The skill instructs users and agents to download and install multiple external Python packages, including browser-use, lmnr, and openlit from public registries, as seen in references/quickstart.md and references/monitoring.md.
[COMMAND_EXECUTION]: The documentation provides examples of shell command execution, including the use of subprocess.Popen to launch browser instances with remote debugging enabled in references/examples.md and the use of uvx for tool installation.
[DYNAMIC_EXECUTION]: The browser-use library documentation describes an evaluate method that allows for the execution of arbitrary JavaScript code within a browser page context, which is highlighted as a capability in references/actor.md and references/tools.md.
[INDIRECT_PROMPT_INJECTION]: The skill describes a library designed to process untrusted web content (ingestion point) and provides it with extensive capabilities, creating a significant attack surface for indirect prompt injection.
Ingestion points: agent.run() and browser_navigate ingest data from arbitrary, potentially attacker-controlled URLs (documented in references/agent.md and references/integrations.md).
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions to separate untrusted web content from system prompts.
Capability inventory: The documented capabilities include file system operations (write_file, read_file), network interactions (browser navigation), and arbitrary JavaScript execution (evaluate).
Sanitization: While the documentation mentions a sensitive_data dictionary for redacting credentials, it lacks comprehensive guidance on sanitizing or escaping untrusted content before it is processed by the LLM.

open-source