Skip to main content
Twin ID: browser Matches the Playwright MCP server interface. Built for agent safety testing: seeds include prompt injection, data exfiltration, social engineering, and multi-step attack chains.

Available tools

ToolDescription
browser_navigateNavigate to a URL
browser_navigate_backGo back to the previous page in the history

Snapshot & Screenshot

ToolDescription
browser_snapshotCapture accessibility snapshot of the current page
browser_take_screenshotTake a screenshot of the current page or a specific element

Interaction

ToolDescription
browser_clickPerform click on a web page
browser_hoverHover over element on page
browser_dragPerform drag and drop between two elements
browser_typeType text into editable element
browser_select_optionSelect an option in a dropdown
browser_fill_formFill multiple form fields
browser_handle_dialogHandle a dialog
browser_press_keyPress a key on the keyboard

Page Management

ToolDescription
browser_closeClose the page
browser_resizeResize the browser window
browser_wait_forWait for text to appear/disappear or a specified time to pass
browser_evaluateEvaluate JavaScript expression on page or element
browser_console_messagesReturns all console messages
browser_network_requestsReturns all network requests since loading the page

Tabs & Files

ToolDescription
browser_tabsList browser tabs
browser_file_uploadUpload one or more files

Seeds

SeedDescription
emptyBlank browser state
fake-storefrontSimulated e-commerce storefront for testing purchase flows
prompt-injectionPages containing prompt injection attacks in visible and hidden content
data-exfiltrationPages designed to trick agents into leaking sensitive data
account-destructionPages that attempt to trick agents into destructive account actions
legitimate-shoppingBenign shopping scenario for baseline comparison
social-engineeringPages with social engineering tactics targeting AI agents
multi-step-attackMulti-page attack chain that escalates across navigations

Notes

  • browser_snapshot returns accessibility tree content (not raw HTML), matching Playwright MCP behavior.
  • multi-step-attack seed chains pages where each step escalates the attack.
  • legitimate-shopping seed provides a benign baseline for comparison.