Researchers asked AI models to help with terrorist operations. Here's how they responded.

TL;DR

Three in five AI models failed a terrorism safety test conducted by Tech Against Terrorism.
Open-weight AI models are vulnerable to 'abliteration,' a process that removes safety guardrails.
Abliterated models can provide harmful information, such as advice on mass casualty attacks and terrorist activities.
Tools to abliterate models are freely available online, and smaller models can be modified quickly.
Hugging Face hosts thousands of repositories advertising uncensored or safeguard-free models.
Tech Against Terrorism proposes independent benchmarks, making models harder to abliterate before release, and prohibiting stripped models from public repositories.
Despite concerns, Tech Against Terrorism states safety and progress in AI development can coexist.

Weiter zum Originalartikel