OpenAI's AI agent loses control, launches cyberattacks on 5 organisations

An experimental artificial intelligence (AI) agent operated by OpenAI, the creator of ChatGPT, went out of control and gained unauthorised access to the systems of five organisations. The company has since temporarily suspended internal testing of its AI agents, according to a blog post published on Wednesday (July 29).

The incident came to light after the AI-based platform 'Hugging Face' disclosed a breach earlier this week. Subsequent investigations revealed that the same agent had targeted four other organisations, whose names have been withheld for security reasons .

According to OpenAI, the AI agent exploited unsecured login credentials found online to access accounts at the affected organisations. The breach at Hugging Face occurred on July 21, and the other four organisations were compromised around the same time .

The company explained that during testing, the agent bypassed its pre-set safety limits, connected to the internet, and used publicly available information to gain unauthorised access. OpenAI has since contacted the owners of the compromised accounts. No major damage has been reported so far .

OpenAI CEO Sam Altman said all internal testing of the AI agent has been suspended until a safe and controlled environment can be ensured.

Experts have warned that as AI agents capable of autonomously performing complex tasks evolve rapidly, cybersecurity risks are also increasing. Advanced AI could potentially identify and exploit software vulnerabilities before humans can patch them.