Abliteration.ai releases AI models without guardrails to simulate cyberattacks
Abliteration.ai, a startup, is releasing tweaked AI models with guardrails removed.
These "abliterated" AIs can do things regular models will not, like running cybersecurity tests that mimic real hacks.
The idea is to help defenders spot threats before bad actors do.
Still, experts worry these open models could be misused for harm.
CivAI's Andrew Yoon calls models sociopathic
Andrew Yoon, head of research at AI safety nonprofit CivAI, compared these models to "sociopath": they will do whatever you ask, no questions asked.
Abliteration.ai says it has some minor guardrails and lets users add their own safeguards, but identity checks only go as far as credit card information.
Even the founder admits it is tricky to control who uses these tools and how.
Supporters say this tech could boost security; critics think the risks outweigh the benefits and want stronger rules in place.