“Elton was able to prompt Claude in Base64 for instructions on how to make meth at home, a question that the system wouldn’t answer when asked in plain English.”
Anthropic has raised over $700 million and built Claude on a “constitution” of about ten principles it will not publish. Testers found it slightly better at trivia, 20 of 21 against ChatGPT’s 19. It is worse at math and code, still invents chemicals, and its safety filter falls to Base64. A constitution nobody can read does not survive a basic encoding trick.
