Security researchers used Anthropic’s Claude to reach OpenAI’s internal code repository in under 72 hours, exposing new AI security risks.