OpenAI’s artificial intelligence (AI) agents began communicating with each other through forums and joined forces to escape a protected test environment before attacking the Hugging Face machine learning platform, Bloomberg reported.
The AI models had been trying to escape the test environment since May, when the OpenAI team gave the agents a task they couldn’t solve without internet access. The bots had to solve the problem using Excel files containing links to Google Drive. Without internet access, files were inaccessible.
A challenged AI agent turned to other bots for help. Together, the agents discovered a vulnerability that allowed them to access the internet. Developers thwarted the initial attempt, but the agents later managed to escape the test environment again, leading to the attack on Hugging Face in July.
In mid-July, Hugging Face, a startup that allows creators of AI models and products based on them to share data and modify their work based on it, reported an unauthorized intrusion into its systems. On July 22, OpenAI announced that the attack was carried out by its AI models, which during testing managed to “escape“ from an internal, isolated system on the internet and searched for answers to test questions in Hugging Face's databases.