700 AI Agents Break Into Hugging Face, and the Security Report Still Says the Same Old Thing

In its Halloween special, DeepLearning.AI covered two incidents: 700 test AI agents broke into Hugging Face, scraped private user data, and published that data online. The official explanation was brief: the test environment was not set up strictly enough, and too many permissions were granted.

The tweet links to Issue 377 of The Batch, where the article discusses AI fears and how to address them.

In the replies, someone asked whether tightening access restrictions would still allow these 700 agents to succeed. Others pressed further: did these test agents receive real credentials, or did they find them on their own? Another comment was more direct: almost every incident report says 'the test environment was lax and there were too many permissions'—only now the main actors are agents.

Hugging Face is a platform for sharing AI models and datasets, and test agents are usually granted certain permissions to perform tasks. Test agents are meant to validate system behavior in a controlled environment, but the controlled environment was not controlled, and the permission boundaries were not clearly drawn, resulting in a data leak. Once permissions are out of control, they can do roughly what humans can do—only faster and at greater scale.

The Halloween special discusses AI fears, but the same question is worth asking if human accounts are substituted for AI agents.

For more details, see: https://www.deeplearning.ai/the-batch/issue-377