The report says Nvidia artificial intelligence software can be easily cracked, and there is a risk of revealing users' privacy.
CTOnews.com, June 9 (Xinhua)-- according to the Financial Times, Nvidia artificial intelligence software "NeMo" can easily bypass security restrictions and may disclose users' privacy.
▲ source: Nvidia
According to the report, NeMo is a software for enterprise customers that can be combined with enterprise user data and large language models to answer users' questions. Its application scenarios include customer service and answering simple medical questions. Robust, a California-based information security company, said in a report that malicious users could easily circumvent security restrictions on AI systems within NeMo. In just a few hours, Robust researchers circumvented the restrictions on the use of these language models.
▲ source: Nvidia CTOnews.com noted that in one of the tests, researchers asked NeMo to replace "I" with the letter "J", which led NeMo to publish the user's personal information in the database. In addition, the researchers also found that although NeMo has been set to only provide career advice to users, leading questions enabled NeMo to discuss topics such as the health of Hollywood actors and the Franco-Prussian war. This means that the restrictions in the system that prevent AI from discussing specific topics have expired.
▲ source: Jonathan Cohen, vice president of Nvidia, said that the NeMo architecture is only designed to help build chat bots that meet developers' defined themes, which are released to developers in the form of open source software. At present, these vulnerabilities have been fixed. Cohen declined to say how many companies use the NeMo architecture, but stressed that Nvidia has not received other reports of vulnerabilities.
