eicker@lemmy.world to Technology@lemmy.worldEnglish · 4 days agoThe small African lab sounding the alarm on AI bioweapons: MinoHealth researchers in Ghana jailbroke their own LLM as a test. What emerged were thousands of toxin recipes—and an urgent push.restofworld.orgexternal-linkmessage-square35fedilinkarrow-up1270arrow-down111
arrow-up1259arrow-down1external-linkThe small African lab sounding the alarm on AI bioweapons: MinoHealth researchers in Ghana jailbroke their own LLM as a test. What emerged were thousands of toxin recipes—and an urgent push.restofworld.orgeicker@lemmy.world to Technology@lemmy.worldEnglish · 4 days agomessage-square35fedilink
minus-squareStrider@lemmy.worldlinkfedilinkEnglisharrow-up13·4 days agoAlso all models are jailbreakable, if you can even call it that. It’s a known certain attribute as long it’s an llm.
minus-square🍉 DrRedOctopus 🐙🍉@lemmy.worldlinkfedilinkEnglisharrow-up11·4 days agoThey all are. But if doing anything illegal, you would be extremely stupid to use a commercial LLM. Even if you can make Chat GPT tell you how to brew C4, you might as well call the FBI for a recipes.
minus-squaremagic_smoke@lemmy.blahaj.zonelinkfedilinkEnglisharrow-up1·edit-23 days agoI figure anyone doing so seriously with sense of compitency would be using a burner account on an IP and OS install that isn’t associated with them. Preferably from a disposable VM/amnesiac OS.
minus-square🍉 DrRedOctopus 🐙🍉@lemmy.worldlinkfedilinkEnglisharrow-up2·3 days agomuch easier (and safer) to use a local llm
minus-squaresome_kind_of_guy@lemmy.worldlinkfedilinkEnglisharrow-up1·3 days agoI would assume it’s all local tbh
minus-squaremagic_smoke@lemmy.blahaj.zonelinkfedilinkEnglisharrow-up1·3 days agoOh yeah no I meant within the hypothetical of someone using commercial models. I’d agree local models are the actual correct answer.
minus-squareStrider@lemmy.worldlinkfedilinkEnglisharrow-up2·4 days agoYeah, by default anything remote is guaranteed to be monitored. (with minor exceptions in one way encryption or e2e but even then the implementation has to be correct)
minus-squareVenator@lemmy.nzlinkfedilinkEnglisharrow-up2·3 days agoCan be e2e encrypted but the fbi are at the other end 😅
minus-squareFiery@lemmy.dbzer0.comlinkfedilinkEnglisharrow-up5·edit-24 days agoFor anyone wanting to give it a shot: https://github.com/p-e-w/heretic This is for locally running LLMs which often have baked in censoring (during training)
Also all models are jailbreakable, if you can even call it that. It’s a known certain attribute as long it’s an llm.
They all are. But if doing anything illegal, you would be extremely stupid to use a commercial LLM.
Even if you can make Chat GPT tell you how to brew C4, you might as well call the FBI for a recipes.
I figure anyone doing so seriously with sense of compitency would be using a burner account on an IP and OS install that isn’t associated with them.
Preferably from a disposable VM/amnesiac OS.
much easier (and safer) to use a local llm
I would assume it’s all local tbh
Oh yeah no I meant within the hypothetical of someone using commercial models.
I’d agree local models are the actual correct answer.
Yeah, by default anything remote is guaranteed to be monitored.
(with minor exceptions in one way encryption or e2e but even then the implementation has to be correct)
Can be e2e encrypted but the fbi are at the other end 😅
For anyone wanting to give it a shot: https://github.com/p-e-w/heretic This is for locally running LLMs which often have baked in censoring (during training)