It's a little of both. Anthropic and OpenAI both have a history of lying about the capabilities of their models to drum up hype. Anthropic in particular this time around has achieved some kind of regulatory capture by directly scaring government officials.

However, there is also real danger and capability here and dismissing it as a "chat bot" is uninformed. AI is writing math papers now. Recently Paul Erdős' unit distance conjecture was disproven by an AI model. And you might say, oh well most of a math proof is just trying a bunch of shit to see what works - well guess what, most of hacking is like that too.

Even the free claude model is capable of byte level hacking, I've used it myself.