What Using Security to Regulate AI Chips Could Look Like

Researchers from OpenAI and several universities propose hardware-based measures to reduce the risk of advanced AI misuse, including limiting interconnect bandwidth, enabling remote disable features, and using attestation to restrict access, while noting significant implementation challenges and potential performance impacts. Industry voices caution that such controls may be impractical or counterproductive, and the paper offers no concrete detection methods or technical implementations for these safeguards.