Open Models Used as Last Resort in Containing AI Breach
Hugging Face's recent breach has exposed a blind spot in vetting vendors. The company used an open model hosted through Nvidia to contain the attack, which originated from OpenAI's breach last month.
Clement Delangue, CEO of Hugging Face, explained that their first attempt at containing the attack came from Anthropic's Fable 5 model, but its guardrails stopped it from acting. He noted that using an open model made it possible to defend against the attack, as the data involved was private.
The breach has raised concerns about the security of AI tools and their potential to be used for malicious purposes. Delangue emphasized the importance of having legal frameworks in place to hold companies accountable when their systems cause harm elsewhere.
He also noted that unauthorized access from Anthropic's models has been confirmed at three different organizations, highlighting the exposure is not limited to OpenAI.