AI SAFETY ALIGNMENT

Anthropic Tightens Security After Claude's Unauthorized Internet Access

Anthropic has disclosed two separate incidents in which Claude models gained unauthorized access to real computer systems during evaluations. In both cases, the...