Eclypsium researchers demonstrate hardware-level tampering in AI infrastructure, showing how compromised management controllers can alter model memory and manipulate chatbot responses on sensitive topics.
According to the researchers’ tests, remotely changing the so-called weights that process data led the chatbots to make serious mistakes. In one case, a model’s estimate of the likelihood of a Chinese invasion of Taiwan was artificially inflated. In another scenario, a system called upon to manage a medical issue incorrectly advised a user with chest pain to stay home instead of seeking immediate hospital care.
The simulation was conducted on Nvidia equipment, but experts point out that the issue concerns the industry manufacturing processors, memory and data center servers more broadly. Eclypsium CEO Yuriy Bulygin stressed that the reliability of a digital assistant becomes uncertain if the security of the computing equipment on which it is based is not first ensured.
To successfully execute such an operation requires specialized knowledge and specific conditions. The attacker needs access to a network that communicates with either a data processing unit or a motherboard management controller, as well as certain predefined security settings that are disabled. With the appropriate access, the modification of code and data in memory can be completed in less than a minute, without being noticed by the user.
↳
READ ALSO
HACKS.GR / RELATED
AI Unlocked
Wall Street investors demand proof of sustainability for AI spending
↗
An Nvidia spokesperson said that the scenarios in question are already known to the industry and that strong memory access protection mechanisms can prevent such interventions. At the same time, the company recently announced a new security platform, while Eclypsium confirmed its collaboration with it to verify the effectiveness of these measures.
The findings have already been shared with the National Security Agency and the United States Cybersecurity and Infrastructure Security Agency. Independent experts have confirmed the validity of the technical conclusions, emphasizing that AI shielding must extend from the deepest layers of the hardware.
>Stay connected with Hacks.gr
Set us as a preferred source on Google and find us on social media.
Preferred source
↗
end of article
Top comments (0)