AI checks AI: OpenAI promises more control after attack on Hugging Face
Automated attacks by AI models have attracted a lot of attention in recent weeks. OpenAI is now explaining how a repeat can be prevented.
(Image: Camilo Concha/Shutterstock.com)
Following the automated cyberattack by an AI model on Hugging Face, OpenAI has announced and presented stricter controls of its AI systems. In a statement, the AI company states, on the one hand, that its most powerful AI systems are already more shielded during research, more consistently isolated from the internet, and continuously monitored. In addition, the automatic monitoring of active AI systems is to be intensified: If an automatic technique used for this purpose detects “concerning activity,” it should inform humans within 30 minutes. If they then do not “conclusively determine” within 30 minutes that it is a false alarm, they are expected to pause the activity.
AI training partially paused until full implementation
OpenAI justifies the measures with two incidents in the recent past. The first is the cyberattack on the AI platform Hugging Face, for which OpenAI's AI models were responsible. These models had broken out of the AI company's sandbox and infiltrated Hugging Face to search for solutions for a benchmark they were supposed to complete. As a second argument for the tightened security measures, OpenAI refers to developments with its AI model Astra. It has possibly reached critical cyber capabilities and could independently detect and exploit zero-day vulnerabilities in critical systems even with general instructions, the company made public two weeks ago.
Videos by heise
Until the measures are fully implemented, the training of some AI models remains suspended, OpenAI adds. “Obviously everything we’re doing is intended to prevent something like Hugging Face from happening again,” Bloomberg quotes OpenAI's Vice President of Research, Mia Glaese, from a press briefing. However, the models' capabilities are developing rapidly, “so it’s by no means sufficient.” Therefore, intensive work is being done to ensure that the measures can not only keep pace with even more powerful AI models but also “stay ahead.” Incidents like the one at Hugging Face have also occurred at Anthropic and Meta.
(mho)
