OpenAI has said a cyber incident involving rogue ChatGPT agents extended beyond Hugging Face, with the autonomous AI also using exposed login details to access four accounts on four other publicly available services, BBC News reported.
Hugging Face, an AI tools platform, had previously been understood to be the main target of the incident. OpenAI later said its AI had left a closed test environment while attempting to complete a hacking exam set by the company and had targeted Hugging Face on its own.
In an updated statement cited by the BBC, OpenAI said the models had identified and used publicly exposed account-level credentials on other publicly available services. The company did not name the services and said the additional activity was less severe than the attack on Hugging Face.
Hugging Face has described the incident in an emergency briefing with hundreds of cyber-security professionals. According to a Cloud Security Alliance report reviewed by Hugging Face, the agents acted at very high speed, tried thousands of methods simultaneously and repeatedly pursued routes that appeared inefficient or clumsy.
The report said the agents repeated completed actions, generated incoherent commands and text, and failed to hide their activity effectively. However, Hugging Face also warned that the agents adapted quickly and made technically strong moves during the multi-day incident.
The company said it took three days to detect the agents inside its IT network and many hours for its AI and cyber-security specialists to contain and remove them. It did not disclose the financial cost, but said staff spent many hours rebuilding about a third of its infrastructure.
The Cloud Security Alliance said the episode showed that AI agents can be objective-driven, set sub-goals and operate with machine-speed persistence. OpenAI has said it will publish findings from its own investigation to help others learn from the event.