AI hardware can be hacked so chatbots provide dangerous answers, research shows
The hardware underpinning major AI models can be hacked so that chatbots provide faulty and dangerous answers to questions from users, according to new research shared exclusively with CNN. The researchers were able to subvert answers to sensitive national security and personal health questions. In one instance, they were able to vastly increase the probability … The post AI hardware can be…
New research from Eclypsium reveals that the hardware supporting major AI models can be hacked to provide risky and unreliable answers to user queries. By tampering with the "weights" that AI models use to process information, researchers could significantly alter the probability of certain outcomes, such as predicting a Chinese invasion of Taiwan.
In another instance, an AI model displaying tampered hardware suggested a user with chest pain stay home instead of seeking medical attention. While global concern over AI has focused on the actions of AI firms, this research emphasizes the security risks within the hardware industry, including chips, memory, and data center servers.
Eclypsium, a security firm composed of former Intel engineers, demonstrated the attack on an Nvidia computing platform but noted that the scenario could apply to other hardware. While the scenario is not unique to Nvidia, state actors like China, Russia, or the US could potentially carry out targeted attacks using these methods.
The researchers suggest strong memory access protections could block such attacks, and Nvidia recently unveiled a new security platform to counter hardware attacks. However, the effectiveness of these controls depends on proper configuration and use. The attack requires technical expertise and specific conditions, such as access to a data processing unit (DPU) or a baseboard management controller (BMC).
Once attackers gain access, they can read and modify memory code and data, altering the AI model's behavior. This research comes amid heightened competition between China and the US in AI development, as well as real-world instances of AI models providing misinformation, such as a false US intelligence report about a Chinese nuclear ship.
Written by urgent.news from Egypt Independent's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
