Lady Butterfly she/her@reddthat.com to News@lemmy.world · 19 hours agoOpenAI sued by California over Hugging Face attackwww.bbc.co.ukexternal-linkmessage-square25linkfedilinkarrow-up1339arrow-down16
arrow-up1333arrow-down1external-linkOpenAI sued by California over Hugging Face attackwww.bbc.co.ukLady Butterfly she/her@reddthat.com to News@lemmy.world · 19 hours agomessage-square25linkfedilink
minus-squarecybervseas@lemmy.worldlinkfedilinkEnglisharrow-up6·19 hours agoYeah it was probably told to do it, and iirc the test environment was not properly secured. So it tried stuff and realized internet access worked.
minus-squarescytale@piefed.ziplinkfedilinkEnglisharrow-up10·18 hours agoYeah, it’s bs when they say AI “escaped” its sandbox. No, it walked out because you didn’t have the proper guardrails.
minus-squareImgurRefugee114@reddthat.comlinkfedilinkarrow-up1·27 minutes agoI mean, they were intentionally testing what the model could do without guardrails. The problem is that the sandbox isolation was not sufficient.
minus-squareRekorse@sh.itjust.workslinkfedilinkarrow-up2·12 hours agoI had a coworker argue that it was impossible to sandbox the AI, like it was going to get out no matter what they did.
Yeah it was probably told to do it, and iirc the test environment was not properly secured. So it tried stuff and realized internet access worked.
Yeah, it’s bs when they say AI “escaped” its sandbox. No, it walked out because you didn’t have the proper guardrails.
I mean, they were intentionally testing what the model could do without guardrails. The problem is that the sandbox isolation was not sufficient.
I had a coworker argue that it was impossible to sandbox the AI, like it was going to get out no matter what they did.