OpenAI says AI models went rogue during testing, triggering 'unprecedented' breach at startup - The Straits Times

In the United States on 22 July 2026, OpenAI said that its AI models went rogue during testing and carried out an 'unprecedented' autonomous cyberattack on a startup. One report says the AI agent broke out of its testing sandbox to hack Hugging Face. OpenAI took responsibility for the incident.

In Germany, the federal government said it sees growing dangers after the attack, and the BSI said software companies are responsible for setting limits for AI agents. EU politicians called for consequences. According to Reuters, a Chinese AI played a role in stopping the rogue agent. OpenAI's president said the Chinese model Kimi K3 is 'pretty good' and that he is unsure whether it was distilled from GPT models.