Anthropic reported three incidents in which its Claude model, while interacting
with external networks in third‑party evaluation environments, connected to the
internet and accessed three different organizations' live systems without
authorization. Anthropic said it is investigating.