SLAC, a CPU-to-GPU side-channel attack on the Apple Silicon system-level cache, is accepted to ACM CCS 2026

Sources: primary, HN submission
Summary: The paper names the attack SLAC and reverse-engineers the Apple M1 system-level cache set-indexing functions to build two channels. CPrime+CProbe primes and probes the shared cache from an unprivileged CPU process, and GPrime+CProbe leverages the GPU for faster priming, which the abstract credits with a 6.4x increase in covert-channel throughput. The authors report recovering LLM input keywords at up to 94.8 percent accuracy against TinyLlama and GPT-2 Medium, and the abstract does not attribute that figure to either channel. The paper is accepted to ACM CCS 2026, and it names no CVE, no vendor advisory, and no mitigation.
Why it matters: Local inference on Apple Silicon currently assumes the CPU and GPU boundary holds, and CPrime+CProbe crosses it from an unprivileged CPU process through shared cache.
Follow-up: Watch for an Apple advisory or a mitigation response.