In a report, OpenAI has outlined ​two incidents in which AI agents hacked the company's ​own infrastructure.