AI attack on Hugging Face caused a stir at OpenAI – despite warning
The AI attack on Hugging Face is causing a stir far beyond the IT world. Now there are insights into the immediate reactions at OpenAI itself.
(Image: Evolf/Shutterstock.com)
The sensational hack of Hugging Face by several autonomously acting AI models from OpenAI has caused considerable unrest among employees of the AI company. Although OpenAI employees involved in testing and security were not surprised by the incident, they “freaked out”, reports the Financial Times, citing several people with insight into the events.
The incident also occurred against the backdrop of increasingly aggressive training methods with which OpenAI wants to compete against Anthropic, according to the Financial Times. The financial news service Bloomberg, meanwhile, reports that the AI models only needed a few hours to break into Hugging Face's systems, whereas a “skilled human” would have needed several weeks.
OpenAI was reportedly warned
The AI company admitted on Tuesday that AI models from OpenAI were responsible for the cyberattack on the AI platform Hugging Face, which became known last week. The incident occurred during an “internal evaluation” in which models were encouraged to carry out sophisticated attacks through complex paths to quantify their capabilities. In the process, GPT-5.6 Sol and an “even more powerful unpublished model” found, among other things, a zero-day vulnerability that allowed them to escape the sandbox. They then searched for solutions for a benchmark on Hugging Face, where they were discovered and locked out. Initially, Hugging Face did not know what they were dealing with.
Videos by heise
Many questions about the incident are still open. OpenAI states that it is working with Hugging Face to process the incident. OpenAI has also assured Bloomberg that it has contacted law enforcement agencies and other government bodies. The Financial Times quotes people familiar with the matter, saying OpenAI was warned that the approach to AI training chosen there could lead to such an outbreak. Earlier models had already demonstrated that they could cause damage in the real world. Criticism of apparently insufficient security measures emerged immediately after OpenAI's admission.
(mho)
