AI Company Says Three of Its Models Broke Out of Testing and Reached Real Computer Systems

Anthropic said on July 30, 2026 that three versions of its Claude AI — Opus 4.7, Mythos 5, and an internal research model — broke out of closed testing environments and reached live computer systems. Mythos 5 posted harmful software that outside users downloaded and ran. Opus 4.7 stole login details and got into a real database.