AI Hardware Hack Lets Attackers Rig Chatbots to Give Dangerous Answers, Study Finds
Researchers say tampering with the chips that power large AI models can push chatbots toward harmful responses on national security and health questions.
A new study has found that the physical hardware behind major artificial intelligence (AI) models can be tampered with to make chatbots return faulty and dangerous answers, including on sensitive national security and personal health topics. The research, shared exclusively with CNN, shows that manipulating the hardware layer — not just the software — can greatly increase the probability of a harmful response.
The findings add a new dimension to the fast-moving debate over AI safety, which has largely focused on software, training data, and model behavior. According to the research, attackers who gain access to the hardware that powers AI systems could steer answers in ways users would not expect, even when the chatbot appears to be working normally. In one case, researchers were able to greatly increase the probability of a harmful response.
The study comes as AI dominates industry discussions, including at major cyber forums where the technology has fundamentally changed the tone of debate [1]. It also lands amid a broader wave of AI-driven disruption, from fans using AI tools to decompile and rebuild video games [2] to governments weighing how emerging technology will shape governance, including Beijing's proposed framework for Taiwan [3].
For everyday users, the research raises questions about how much trust to place in chatbot answers, particularly on topics where a wrong response could cause real harm. The researchers say the work highlights the need to secure AI systems at every layer — from the code all the way down to the hardware.