Study finds filtered data stops openly-available AI models from performing dangerous tasks

Published on August 11, 2025

Researchers from the University of Oxford, EleutherAI, and the UK AI Security Institute have reported a major advance in safeguarding open-weight language models. By filtering out potentially harmful knowledge during training, the researchers were able to build models that resist subsequent malicious updates – especially valuable in sensitive domains such as biothreat research.
Continue reading on website
Other news

Viking silver hoard reveals far-reaching trade links between England and the Islamic world

August 11, 2025
New research into a remarkable Viking-Age silver hoard discovered in North Yorkshire, led by Dr Jane Kershaw,  Associate Professor of Viking Age Archaeology, School of Archaeology, has shed light on the international scope of Viking wealth - revealing that a significant proportion of the silver originated not from local raids, but from long-distance trade networks extending deep into the Islamic w

Why common blood pressure readings may be misleading – and how to fix them

August 12, 2025
High blood pressure, or hypertension, is the top risk factor for premature death, associated with heart disease, strokes and heart attacks. However, inaccuracies in the most common form of blood pressure measurement mean that as many as 30% of cases of high blood pressure could be missed.The researchers, from the University of Cambridge, built an experimental model that explained the physics behin