AI hardware can be hacked so chatbots provide dangerous answers, research shows

Why it matters

Security firm Eclypsium, in research shared exclusively with CNN, showed that attackers who gain network access to data center hardware can tamper with the weights that AI models use to process information, quietly altering chatbot responses. In demonstrations, tampered models steered a user with chest pain away from seeking hospital care and inflated the likelihood a model assigned to a Chinese invasion of Taiwan. The firm, staffed by engineers formerly at Intel, demonstrated the attack on a computing platform made by Nvidia, which said it found nothing new and that strong memory access protections can block the path; Eclypsium countered that effectiveness depends on configuration. Attackers typically need access to a data processing unit or a baseboard management controller, with certain protections disabled. Eclypsium briefed US security agencies, and independent experts called the findings credible, urging attention to the computing stack beneath model safety debates.