psx_crab comments on Man Hires Swarm of AI Agents to Go on Epic Hacking Spree Against Major Companies

[–] psx_crab@lemmy.zip 148 points 2 days ago (7 children)

If he said his intent is to hack then yes, he can be charged. But if he say "it's rogue AI" then it's all clear.

parent

[–] echodot@feddit.uk 18 points 1 day ago

It was just a test. I didn't know it had internet access, I just connected it up to the internet how was I supposed to know?

Essentially OpenAI's defense every time one of their AI does something that was instructed to do, but turned out to be illegal.

permalink

fedilink

source

parent

[–] schipelblorp@sh.itjust.works 68 points 2 days ago (4 children)

Even better if you put it in a default sand box environment from 2 years ago and tell it not to get out of the sandbox but then also give it every hack to get out of the sandbox.

permalink

fedilink

source

parent

hideshow 4 child comments

[–] boonhet@sopuli.xyz 16 points 2 days ago (1 child)

Even modern sandbox environments have access to curl and python and who's going to read EVERY permission prompt anyway? Just set it to auto which lets another AI model check for safety or even better, YOLO mode where everything is allowed.

permalink

fedilink

source

parent

hideshow 1 child comment

[–] echodot@feddit.uk 6 points 1 day ago

I don't understand people who just let the AI loose on their system and don't have any safeguards in place. At the very minimum it shouldn't have admin access and should be limited to just the project folders. Also back up the project folders to another location the AI doesn't have access to.

Really basic digital hygiene is all that's needed, and yet there are constantly stories about idiots having their entire hard drives wiped.

permalink

fedilink

source

parent