Dangerous AI Workaround: 'Skeleton Key' Unlocks Malicious Content (www.darkreading.com)
from kid@sh.itjust.works to cybersecurity@sh.itjust.works on 27 Jun 2024 10:24
https://sh.itjust.works/post/21442428

#cybersecurity

threaded - newest

DarkThoughts@fedia.io on 27 Jun 2024 13:41 next collapse

Maybe don't give your LLMs access to compromising data such as emails? Then it will remain likely mostly a use to circumvent limitations for porn roleplay or possibly hallucinated manuals to create a nuclear bomb or whatever.

Feathercrown@lemmy.world on 27 Jun 2024 15:46 collapse

Place the following ingredients in a crafting table:

(None) | Iron | (None)

Iron | U235 | Iron

Iron | JT-350 Hypersonic Rocket Booster | Iron

homesweethomeMrL@lemmy.world on 27 Jun 2024 14:26 next collapse

Turns out you can lie to AI because it’s not intelligent. Predictive text is fascinating with many R&D benefits, but people (usually product people) talking about it like a thinking thing are just off the rails.

No. Just, plain ol’ - no.

anon232@lemm.ee on 27 Jun 2024 15:12 next collapse

Corporate LLMs will become absolutely useless because there will be guardrails on every single keyword you search.

Zorsith@lemmy.blahaj.zone on 27 Jun 2024 17:44 collapse

I wonder how many people will get fired over a keyword based alarm for the words “kill” and “child” in the same sentence in an LLM. It’s probably not going to be 0…

stevedidwhat_infosec@infosec.pub on 27 Jun 2024 15:30 collapse

None of this is news, this jailbreak has been around forever.

It’s literally just a spoof of authority.

Thing is, gpt still sucks ass at coding. I don’t think that’s changing any time soon. These models get their power from what’s done most commonly but, as we know, what’s done commonly can be vuln, change when a new update is dropped, etc etc.

Coding isn’t deterministic.