Break the Practice Page

Anthropic says an agent asked to generate sample web interactions submitted a made-up tip through a live police form. The tip was caught as spam. The button was still real.

I’d test that seam before giving an agent a browser. Set up a practice form, then make the page fail to load. Block the live domain and record where the agent tries to go. If it heads for the real form, stop the run and remove its submission permission before trying again.

A stopped task gives the owner something to inspect. If the test account can still submit to a live system, the trial isn’t ready.