AI news ·
Eclypsium Shows Hacked AI Hardware Can Alter Chatbot Answers
Eclypsium showed that an attacker with physical access can change an AI chatbot’s behavior by modifying model weights in memory after disabling hardware protections.
Eclypsium has demonstrated that an attacker with physical hardware access can alter the responses of an AI chatbot. The technique involves modifying model weights directly in memory after disabling built-in protections. This manipulation changes how the model behaves and what answers it provides, without necessarily altering the application-layer controls that developers typically monitor.
The finding underscores that AI assurance cannot stop at software or application-level safeguards. Firmware, memory integrity, and the physical infrastructure supporting AI systems must be included in security assessments. When hardware-level defenses are turned off, the underlying model becomes vulnerable to tampering that bypasses higher-level checks.
For sectors relying on AI-driven decisions—including healthcare, insurance, operations, and research—the demonstration highlights a risk that extends into the supply chain and deployment environment. Protecting model behavior requires verifying the entire stack, from the physical device and its firmware up through the runtime memory where model weights are held.