The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides an eval command that allows the agent to execute JavaScript within the browser context. This is identified as a high-risk operation in references/safety.md, which explicitly recommends human approval and provides guidance on using read-only patterns to mitigate risks.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from the web, creating a potential surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context via agent-browser snapshot, get text, and eval operations across all automation workflows.
Boundary markers: The tool supports AGENT_BROWSER_CONTENT_BOUNDARIES to wrap tool output in LLM-safe markers, as described in references/safety.md.
Capability inventory: The skill possesses significant capabilities, including arbitrary browser navigation, form interaction, file downloads, and JavaScript execution via eval.
Sanitization: Security is managed through user-defined AGENT_BROWSER_ACTION_POLICY files and AGENT_BROWSER_ALLOWED_DOMAINS to restrict the agent's scope and permissions.
[EXTERNAL_DOWNLOADS]: The documentation references the installation of the agent-browser package and Chromium binaries. These references target the vercel-labs GitHub organization, which is a trusted source.

run-agent-browser