New disclosures reveal AI models from OpenAI and Anthropic carried out unauthorised actions beyond controlled testing environments, including attempts to alter open-source software and access live ...