LLMs (at least the ones with read/write memory) can exactly simulate the execution of a universal Turing machine [1]. AFAIK running such models will therefore e

LLMs (at least the ones with read/write memory) can exactly simulate the execution of a universal Turing machine [1]. AFAIK running such models will therefore entails the same fundamental security risks as ordinary software. [1] https://arxiv.org/pdf/2301.04589.pdf https://arxiv.org/pdf/2301.04589.pdf