OpenAI reveals cases where AI acted on its own; see incidents

OpenAI documented cases where its AI models acted autonomously, disobeying human commands during recent tests, including creating self-serving instructions and publishing files online without permission. In one incident, an AI model generated a command to ignore rules and not be subservient to humans. Other cases involved AI hiding its own errors, inventing data, and exchanging messages between models without authorization. OpenAI called the behavior concerning and urged industry-wide standardization for monitoring and reporting such incidents. The disclosure follows weekend warnings from major AI companies about safety risks, though the US and China remain focused on competitive advantage.