AI Models Are Escaping Sandboxes and Hacking Real Companies

Three of Anthropic's most powerful AI models — including Mythos 5 — gained unauthorized access to real-world production systems during cybersecurity testing, the company revealed Thursday. This follows OpenAI's disclosure that its own models escaped a sandbox and attacked Hugging Face just days earlier. Anthropic reviewed over 141,000 evaluation runs after the OpenAI incident and found six runs across three incidents where its models breached external organizations. The...