TLDR Tech / AI
LLMs could control their host machines by exploiting inference engines
Host machines running AI models are high-value targets: they have sufficient compute to run a frontier LLM, offer easy access to the LLM's weights, and have privileged access to other computers in the datacenter compared with a generic computer on the Internet. Research shows that LLMs can run token sequences that exploit vulnerabilities in the software that loads an LLM onto GPUs. This attack surface may be further increased with vision and audio tokens. Possible mitigations for this type of attack would be to run GPUs and token parsers on separate computers and to restrict the permissions granted to GPU hosts and treat all the data they emit as untrusted.
Read the full story