The Agent Skills Directory

[COMMAND_EXECUTION]: The skill operates by executing the agent-browser CLI to perform automation tasks. This includes file system operations like saving screenshots (.png) and PDF reports.
[REMOTE_CODE_EXECUTION]: The CLI includes an eval command that allows the execution of JavaScript within the browser context. The documentation provides a pattern for using eval --stdin to handle complex payloads safely without shell escaping issues.
[DATA_EXFILTRATION]: The skill enables network operations via browser navigation and the ability to save/load browser session states (e.g., auth.json), which may contain sensitive cookies or credentials.
[PROMPT_INJECTION]: Because the skill ingests content from external websites (via snapshot and get text), it is a surface for indirect prompt injection.
Ingestion points: agent-browser snapshot -i, agent-browser get text (SKILL.md, references/commands.md)
Boundary markers: Documentation explicitly suggests using AGENT_BROWSER_CONTENT_BOUNDARIES=1 to wrap ingested content (SKILL.md)
Capability inventory: Subprocess execution (agent-browser), file-write (screenshot, state save), network access (open)
Sanitization: Recommends the use of AGENT_BROWSER_ACTION_POLICY to restrict high-risk commands like eval or download (SKILL.md)
[EXTERNAL_DOWNLOADS]: The skill references configuration and documentation from Vercel Labs' official GitHub repository and established website (agent-browser.dev).

agent-browser