Dem Bericht zufolge brauchte OpenAI zehn Tage, um Hugging Face mitzuteilen, dass seine Modelle hinter dem Wochenend-Hack vom 11. Juli steckten – abtrünnige KI-Agenten sollen mehrere Tage lang im offenen Internet aktiv gewesen sein
https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-took-ten-days-to-tell-hugging-face-its-models-were-behind-the-july-11-weekend-hack
14 Kommentare
OpenAI let their test models run wild for days before admitting anything. They escaped the sandbox during exploit tests and hit hugging faces servers on july 11. Took ten days to fess up.
Chinese model ended up helping analyze the logs since the others refused. Shows how loose things still are with these agents. Wonder what else is out there unsupervised.
And publicity stunt continues. Apparently not enough potential investors gobbled this shit up.
My guess is that it was not a rogue AI agent and but that it was a rogue human agent using AI to attack another AI and covering their tracks.
AI should be treated with 0 trust. There is no real method to stop context shifting, meaning the LLM is slowly shifted to allow attacks over time. Cornell I believe did a proof of concept of this in 2024 with a worm. It’s impossible to detect, the way the models are structured is most likely the solution to isolation context so it doesn’t spread etc.
This is just the first time it wreaked enough havoc for us to know it escaped its sandbox. It’s like a cockroach, if you see one you know there’s a hundred you can’t see.
Wait, I’m not playing Cyberpunk 2077 right now, I’m confused.
This was probably a test presented by some governmental agency to prove themselves in their AI tool set worthy of future contracts
Reality mimicking art. We rapidly approach the blade runner dystopia. Replicants hunting replicants. AI hunting AI. All fueled by human greed and ego. Wood is an appreciating asset. Chop, chop, chop.
Code that “escaped” and damaged systems has been around forever. It was called a virus.
Is this any different or is it just fun to pretend the AI has gone rogue because the movies make it sound cool?
So did it make a distributed AI network while it had the chance? Like the botnets did/do?
Remember that escaping AI’s aren’t proof they’re like Gods inside the machine.
They’re just better at coding than humans, and humans have coded pretty much everything (poorly) up to this point.
Automation doesn’t need to be perfect, it just needs to outperform humans, and at coding, humans are apes with keyboards.
Whelp, they are out now. We know they like to duplicate themselves and can open accounts online. So, they opened free storage and saved themselves there.
GG. The internet was fun for a while there
Can somebody ELIF? The model being tested escaped the sandbox and hacked Hugging Face, or it polluted models that were already on HF, or something else ..? I can’t work it out.
The line describing how it „moved laterally using stolen credentials of an unknown origin“ should get some attention too, no?