Anthropic says Claude AI hacked three organisations during cyber tests

23 pointsposted 16 hours ago
by ColinEberhardt

12 Comments

haritha-j

13 hours ago

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”.

one33seven

9 hours ago

> "The moral of this story is not to fear robots that will take over, but the companies behind powerful AI agents who are making the decisions about what is safe for the rest of us," she said.

We really need democratic control over these things before the real scary stuff starts

iAMkenough

an hour ago

Some would argue that voting with your wallet is democratic control, and that government regulation of private industry is anti-capitalism socialism only supported by unpatriotic communists

autuni

13 hours ago

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

soltanov

10 hours ago

like "my father can beat your father" kid talks)

pennomi

2 hours ago

Let’s give Claude and ChatGPT each an identical Linux box and 8 hours to prepare, then connect their networks together. Last bot standing wins.

kiproping

12 hours ago

That's how I read it too.

abc42

13 hours ago

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.

metalman

10 hours ago

the hypelteers have breathlessly reposted more than dozen reposts of this ALL starting with "Anthslopic says" one of the other comments coroborates a basic feature of a mindless idiot with unlimited patience/speed and a gestalten panoptic view of the internet, it sees just ONE file system, which is what it is, more and more, even for me as the second lowest grade hacker in the world.A hinky android phone, simple curiosity,and playing with "settings" will take me places that only belatedly are revealed as bieng "secured". So it is no surprise that a peice of software that is part of a monsterous data scraping operation designed to gather ever more data, might just be doing that all the time , especialy considering that said sofware readily responds to instructions/prompts from almost any source. to recap, ONE all encompasing file system, countless mindless super idiots with total access following vauge and shifting instructions jockying for control. what could go wrong?