Tag: AI safety

A new theoretical review argues that reward hacking, benchmark failure, and distribution shift share a single structural root: observation regimes ...

Researchers argue that multi-step agentic AI systems break traditional clinical validation models and propose structured testing harnesses adapted from simulation, ...

A systematic review of 93 studies reveals eight attack types and nine defence types across six categories of unsupervised machine ...

Researchers have developed a hypernetwork-based framework that lets continually learning AI systems erase specific tasks without access to the original ...

A new Nature Machine Intelligence study provides causal evidence that language models actively use internal confidence signals to shape their ...