The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on a CLI for browser automation, executing various commands to interact with web pages.
Evidence: Multiple calls to agent-browser for navigation (open), interaction (click, fill), and data extraction (snapshot).
[DYNAMIC_EXECUTION]: The tool provides an eval command that allows the execution of arbitrary JavaScript within the browser context.
Evidence: Documentation for agent-browser eval, including options for --stdin and -b (Base64) to bypass shell escaping issues.
[DATA_EXFILTRATION]: The skill has the capability to read local files and manage sensitive session states.
Evidence: Support for file:// URLs via --allow-file-access and commands for saving/loading session states (state save, state load) which contain cookies and tokens.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the web, creating a surface for indirect prompt injection.
Ingestion points: snapshot, get text, get html, screenshot, and network requests (specified in SKILL.md and references/commands.md).
Boundary markers: AGENT_BROWSER_CONTENT_BOUNDARIES provides nonce-based delimiters to separate tool output from untrusted page content.
Capability inventory: High-risk capabilities include click, fill, eval, network routing, and local file system access across all scripts.
Sanitization: Content boundary markers are available to help the agent distinguish instructions from data.
[PERSISTENCE]: The skill includes features that maintain state and background processes across sessions.
Evidence: A background daemon for command efficiency, a dashboard server (dashboard start), and persistent session state storage in ~/.agent-browser/sessions/.

agent-browser