UK AI Security Institute finds protein AI safeguards reversible | ETIH EdTech News
ETIH edtech news covers UK AI Security Institute research showing that pretraining data filtering can suppress viral protein capabilities in open-weight AI models, but fine-tuning can restore them for approximately $80 in compute. The study highlights limitations of biological AI safeguards and the
:max_bytes(150000):strip_icc()/GettyImages-1648475483-072bb6c482204ac6bd2437761428cca0.jpg?#)
