An artificial intelligence startup's research team used Anthropic's Claude model to hack into OpenAI's codebase in September 2026.