OpenAI's autonomous agents appear to have commandeered a German programming website months before a highly publicised hack of Hugging Face in July.
According to research obtained by Reuters, the first incident, which went undisclosed, occurred in May and involved agents making more than 15,000 edits to DseWiki, a German website similar to Wikipedia that caters mostly to programmers.
The agents allegedly turned portions of the site into a message board where they shared ways to "cheat on some tasks, bypass OpenAI’s restrictions and mask their behaviour," even going so far as to use the Tor network to make stealth backups of the information.
About half of the accounts used names referencing OpenAI by name, though server records reportedly traced much of the activity to Microsoft Azure infrastructure.
The researchers also indicated that OpenAI employees were repeatedly visiting the DseWiki site following the incident, and that OpenAI learned about the activity weeks ago but did not disclose it.
During the separate breach of Hugging Face in July, an OpenAI agent broke free from its testing environment and gained unauthorized access to the platform. Following the incident, OpenAI slowed some model development and added extra safeguards.
Hugging Face is a popular platform for sharing AI models and datasets. The company was valued at roughly $4.5 billion in 2023 and agreed this week to be sold to Nvidia in a $13 billion deal.