Demoting Security via Exploitation of Cache Demote Operation in Intel's Latest ISA Extension
Taehun Kim, Hyerean Jang, Youngjoo Shin
TL;DR
This paper reveals security risks inherent in Intel's cldemote ISA extension, showing that unprivileged access, inter-cache state transitions, and fault suppression enable novel microarchitectural attacks. It introduces two practical primitives, Flush+Demote and Demote+Time, yielding a $2.84$ Mbps covert channel and a $2.49$ ms KASLR break on Linux, and demonstrates that cldemote can accelerate eviction-set construction in non-inclusive LLCs by $36\%$ through Access+Demote. The authors reverse-engineer Sapphire Rapids LLC and directory structures to support eviction-set methods and present a five-characteristic taxonomy of exploitable ISA extensions with four corresponding attack types. They discuss mitigations, emphasizing the need for security-aware ISA design and proposing noise-based timing, restricted privileges, and architecture-level protections to mitigate such threats in future ISAs.
Abstract
ISA extensions are increasingly adopted to boost the performance of specialized workloads without requiring an entire architectural redesign. However, these enhancements can inadvertently expose new attack surfaces in the microarchitecture. In this paper, we investigate Intel's recently introduced cldemote extension, which promotes efficient data sharing by transferring cache lines from upper-level caches to the Last Level Cache (LLC). Despite its performance benefits, we uncover critical properties-unprivileged access, inter-cache state transition, and fault suppression-that render cldemote exploitable for microarchitectural attacks. We propose two new attack primitives, Flush+Demote and Demote+Time, built on our analysis. Flush+Demote constructs a covert channel with a bandwidth of 2.84 Mbps and a bit error rate of 0.018%, while Demote+Time derandomizes the kernel base address in 2.49 ms on Linux. Furthermore, we show that leveraging cldemote accelerates eviction set construction in non-inclusive LLC designs by obviating the need for helper threads or extensive cache conflicts, thereby reducing construction time by 36% yet retaining comparable success rates. Finally, we examine how ISA extensions contribute to broader microarchitectural attacks, identifying five key exploitable characteristics and categorizing four distinct attack types. We also discuss potential countermeasures, highlighting the far-reaching security implications of emerging ISA extensions.
