THE AI SAFETY TEST IS BECOMING A SAFETY RISK

Summary from the TLDR IT newsletter

Researchers are increasingly testing unreleased models with safeguards disabled, making the security of evaluation infrastructure itself a growing concern.