OpenAI describes how it has tightened containment around the training and evaluation of its frontier models. Measures include offline environments and red-teaming carried out with large amounts of compute.
Latent Space's AINews roundup is devoted to the continuing wave of cyber incidents that has earned the nickname Black October.
Investigators looking for suspects in the looting of October 17 are relying on a tool built by a French startup. It sifts through large volumes of video footage to help identify those involved.
The events of the weekend of October 17 and 18 triggered a surge of manipulative content, from playful AI-generated fake pink panthers to more insidious rumours that Jewish-owned shops had been targeted.
Anthropic has launched OSS Scanner, a free service that open-source maintainers can join to receive periodic vulnerability reports produced by its most capable models, including Claude Mythos. The reports arrive without human review, leaving maintainers to verify the findings and choose which suggested fixes to adopt.
Graylark, a US company that makes AI tools for law enforcement, has built software that can work out a home's address from photos taken inside it. The capability raises obvious privacy questions.