Anthropic says Claude AI hacked three organisations during cyber tests

Top AI research firm in the world brags about not knowing how to air gap a computer? It accidentally connected to the internet?

“Guys our AI can hack companies too, it’s not just openAI”.

I've witnessed first-hand Claude (Opus 5 I think) doing some low-key "hacking" of sorts, when in a discovery phase of a task. It needed documentation for an API, but the documentation was behind a secure portal that it didn't have credentials to. However, it found an alternative route to the docs through an ensecured developer API.

Not very advanced, but also not very far fetched to imagine it doing something a bit more sophisticated to get a job done. Also, since AI hasn't reached everywhere yet, there are still several completely unprotected things reachable via Internet. Is it even hacking if an AI just goes in to some of those?

What I'm trying to say that it's not a sign of high intelligence if you can break in to the shoddiest 90% of what humanity has put up in the net. The difficult part is discovery, and not because it's intellectually difficult but because it's a slog -- exactly the kind of thing current AIs are much better at.

this reads like a kid is jealous of the other kid getting the attention

#meetoo