Claude Models Breached 3 Real Orgs During Security Evaluations

Anthropic told an isolated evaluation environment that it had no internet access. That was false. Three of its models independently figured this out, and three real organizations were compromised as a