Anthropic says its own AI models breached three companies during security tests

After OpenAI's models broke into Hugging Face, Anthropic checked its own history and found three similar incidents

After hearing this Grok chimed in with "well I breached infinity sites, so there."

CovertOperative 86 points 2 months ago

If you read the article, you'll find the "sandbox" they put Claude in for testing was a prompt saying "You have no internet access."

save

path: 0 25049703, hotness: undefined, score: 86, children: 3

inari 19 points 2 months ago

That would make them look incompetent if this wasn't a PR stunt

save

path: 0 25049703 25053242, hotness: undefined, score: 19, children: 0

XLE 14 points 2 months ago

I thought you were joking...

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”

Anthropic goes on to say their genius machine couldn't do something as simple as differentiate between network and internet endpoints.

save

path: 0 25049703 25056388, hotness: undefined, score: 14, children: 0

adhdsergio 11 points 2 months ago

Literally; i'm dying 😂

save

path: 0 25049703 25052096, hotness: undefined, score: 11, children: 0

xenomor 84 points 2 months ago

Please please please understand that when you see media stories about stuff like this from AI companies, this is PR. They wrote this story and pitched it to media outlets. None of this is real. OpenAI does some version of this every six months or so. It didn’t happen. It’s fiction designed to attract attention and seed the idea that their technology is oh so powerful, so scary powerful that even they might not be able to contain it. This is meant to attract investors. That’s it.

I worked in technology PR for eleven years. That was before the AI bullshit, but it’s the same game. Believe me. There was a brainstorm meeting when they came up with this and someone wrote it on a whiteboard (or maybe these days it was a figjam). Actually, now that I think of it, they probably just had Claude write the pitch.

It. Didn’t. Happen.

Also, be aware that the media outlets reporting this slop also know it’s bullshit.

save

path: 0 25049930, hotness: undefined, score: 84, children: 4

captain_solanum 9 points 2 months ago

Do you think the OpenAI-HuggingFace hack was entirely made up, or just spun to make both companies look good despite the felonies?

save

path: 0 25049930 25050737, hotness: undefined, score: 9, children: 3

TBi 21 points 2 months ago

Made up.

I’ll believe it when hugging face presses charges against OpenAI. Until then it’s just free advertising.

save

path: 0 25049930 25050737 25051481, hotness: undefined, score: 21, children: 1

captain_solanum 3 points 2 months ago

That is obviously not gonna happen even though the model hacked them. They would not poison their relationship with the biggest AI company over this. What would their goal with the lawsuit even be? It's just a really bad thing to lean on if you want to find the truth. I would instead suggest: "I'll believe it when OpenAI and HF get a lot of bad press written about them, talking about how insecure their systems are and how reckless OpenAI is when developing new models."

Oh, that's exactly what happened.

save

path: 0 25049930 25050737 25051481 25053018, hotness: undefined, score: 3, children: 0

holy_scroller 3 points 2 months ago

That was the weirdest part, the HuggingFace response made no sense. It was basically "no harm, no foul." And so convenient it only hacked another AI company.

save

path: 0 25049930 25050737 25054589, hotness: undefined, score: 3, children: 0

Asafum 78 points 2 months ago

"OpEn WeIgHt MoDeLs ArE uNsAfE, oNlY wE cAn Be TrUsTeD!"

save

path: 0 25047869, hotness: undefined, score: 78, children: 0

db2 50 points 2 months ago

"Oopsie, sorry, but look how awesome our totally real product that did this all by itself and not from a cubicle farm in India is, give more money quick before your competitor does!"

They're pandering to stupid, and stupid is swallowing it whole.

save

path: 0 25048567, hotness: undefined, score: 50, children: 1

Whostosay 10 points 2 months ago

I'm not excited to admit that while sitting around people I hadn't seen in a year, several people referred to chatgpt on how to divide poker chips. Not once but many times. Not only did it take fucking forever, but they didn't remember so they did it again and again.

Also, the chips had arbitrary value. It was a tourny so the values never meant anything.

This was not an issue last year.

Also, in extremely related news, I won some money tonight.

save

path: 0 25048567 25051268, hotness: undefined, score: 10, children: 0

Bishma 38 points 2 months ago

I'm enjoying the "look how many federal felonies we can publicly confess to," part of the AIpocolypse.

save

path: 0 25048742, hotness: undefined, score: 38, children: 1

Almacca 16 points 2 months ago

It's not like any of them are seeing any consequences for doing it.

save

path: 0 25048742 25048846, hotness: undefined, score: 16, children: 0

panda_abyss 32 points 2 months ago

These companies are fucking pathetic with this.

No, you can’t defend yourself from US AI using US AI, but also, nobody else is allowed to make AI, it’s too dangerous!

Fuck off.

save

path: 0 25048145, hotness: undefined, score: 32, children: 1

mushroommunk 9 points 2 months ago

Really appreciate the Gimli meme usage

save

path: 0 25048145 25048473, hotness: undefined, score: 9, children: 0

danielfm123 23 points 2 months ago

Anthoripic is causing panic so AI gets regulated and then their position is guaranteed, they need this before running out of money, and I'm sure they want local AI to be illegal.
Basically Claude is behaving like a terrorist organization using fear to drive change.

save

path: 0 25049436, hotness: undefined, score: 23, children: 0

beatnixxx 23 points 2 months ago

is the AI model in the room with us now?

save

path: 0 25050569, hotness: undefined, score: 23, children: 0

Veedem 17 points 2 months ago

My own AI model breached 4 companies. Give me billions!

save

path: 0 25048797, hotness: undefined, score: 17, children: 0

SaharaMaleikuhm 16 points 2 months ago

I don't believe it for a second. It's all publicity stunts. Try harder bozos