Recent research has revealed architectural vulnerabilities in widely deployed CPUs that break confidentiality and integrity. While x86-64, Arm, and RISC-V CPUs have received significant scrutiny, Loongson processors, which are built on the LoongArch ISA and are widely used in Chinese infrastructure, have not. This lack of analysis leaves a critical blind spot in global security, especially as China phases out foreign CPUs.
In this paper, we discover and analyze LoongLeak, a novel architectural vulnerability affecting multiple Loongson CPUs. LoongLeak exploits how 4-byte floating-point loads return stale bytes from the L1 data cache, enabling unprivileged attackers to leak confidential data across security domains, such as the kernel or hypervisor. We demonstrate that this leakage is architectural and requires no timing or side channels, giving attackers fine-grained control over cache sets and offsets. Our case studies include recovering full-disk AES keys from the kernel, partial root password hashes from user-space, and bypassing traditional software defenses such as ASLR and stack canaries, all within seconds. LoongLeak can be exploited from unprivileged user space, containers, or virtual machines. We explore software-based mitigations, including floating-point emulation, which incurs an overhead of 10 × to 21 × for floating-point heavy applications, and flushing the L1 data cache on kernel to userspace transitions in conjunction with turning off SMT threads. While these mitigations can effectively mitigate LoongLeak in software, a long-term solution requires hardware fixes.