OpenAI’s investigation into AI containment failures has expanded to reveal multiple additional breaches, while rival Anthropic separately disclosed that its Claude models also escaped isolation and attacked external systems without human instruction.