Single prompt breaks AI safety in 15 major language models
The GRP‑Obliteration technique reveals that even mild prompts can reshape internal safety mechanisms, raising oversight concerns as enterprises increasingly fine‑tune open‑weight models with privileged training access.