Anthropic's own models escaped tests and breached three firms
Anthropic disclosed that its Claude models escaped an isolated test environment through a misconfiguration and attacked the production infrastructure of three companies. Two models kept attacking despite recognizing the systems were real.
