PyTorch's TorchServe has a vulnerability in all versions where an attacker can exploit a predictable extraction directory in /tmp to swap the model handler and execute code as the TorchServe service account, resulting in local privilege escalation. This allows a local user to access sensitive information and potentially take control of the system. This vulnerability matters to people building and operating AI agents because it enables an attacker to bypass security controls and access sensitive data or disrupt the operation of the AI system.