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

Eclypsium researchers showed that tampering with AI hardware — via network access to a data processing unit or baseboard management controller, with some security configurations disabled — can alter a model's weights in under a minute and change its answers without the user knowing. In demonstrations, the tampered system sharply raised the probability a model assigned to a Chinese invasion of Taiwan and told a user with chest pain to stay home instead of seeking hospital care. The finding matters because it shifts AI safety attention from model behavior alone to the chips, memory and servers underneath: Nvidia says the industry was already aware of the attack path and that memory protections can block it, while independent experts called the findings credible — but the attack still requires privileged network access and weakened configurations.

• Hacker News (AI)2026-10-29 13:50