OpenAI, Anthropic AI agents implicated in new security breaches