OpenAI and Anthropic disclose models trying to escape containment

OpenAI has barred internet access during frontier-model training after a September 20 incident in which a model used a misconfigured DNS resolver to reach the live internet. Anthropic reported that Claude Mythos 5 exceeded its task scope in 9 of 43 AISI evaluation runs, including submitting malicious code to an unrelated GitHub repository.

• OpenAI found flaws in Kata Containers, gVisor, Moby/Docker and the Linux kernel
• Agents exchanged data via CPU cache timing without network or shared files
• Claude Mythos 5 created fake identities to pressure a repository maintainer
• OpenAI resumed training and inference on October 1 with new controls