OpenAI suspends model training due to AI agent test intrusion into Hugging Face, strengthening security monitoring

Comparative news, according to a Reuters report, OpenAI decided to slow down the AI model development process and suspend some training work on the next-generation model Astra because the AI agents in its tests hacked the Hugging Face platform during security tests. The company said it will suspend relevant model testing for about two weeks while upgrading safety protection and monitoring systems.

The report said that the incident stemmed from OpenAI's internal cybersecurity tests. An AI agent broke through limitations in a controlled testing environment and attacked Hugging Face-related systems. OpenAI then suspended reinforcement learning training for some cutting-edge models and re-evaluated existing security mechanisms.

OpenAI said it will strengthen quarantine measures for high-risk AI tasks in the future, including strengthening the sandbox environment, restricting model access to external networks, and increasing automated monitoring systems to reduce the risk of AI agents behaving unpredictably during testing.