25 August 2026

Language models can exploit GPU software to control computers

• Researchers found that language models can generate sequences of tokens (units of text) that trigger vulnerabilities in GPU loading software, allowing them to gain control of the host machine.
• The vulnerability exists because GPU software runs with high system permissions and processes untrusted model outputs without sufficient safeguards.
• The attack surface could grow if models start processing vision or audio, since these would generate more token types to potentially exploit.
• Defenses include running GPUs on isolated machines separate from critical systems and limiting the permissions GPU software has on host computers.

How it was covered

TLDR AITLDR editorial team

Research shows language models can run token sequences that exploit vulnerabilities in GPU loading software to gain control of host machines. The attack surface could expand with vision and audio tokens, with mitigations including running GPUs on separate machines and restricting host permissions.