Black Hat Asia 2026 | Revealing User Activity on macOS for Apple Silicon
Researchers have developed Tide, a novel method to detect interrupts on Apple Silicon without relying on timers, enabling the reconstruction of user activity through side-channel analysis. By exploiting deterministic patterns in the XNU kernel's exception handling and the ARM calling convention, specifically the usage of register x18, the study demonstrates that applications running in EL0 can infer keystrokes and other inputs from EL1. This finding challenges the security assumptions of M-series chips, proving that hardware isolation between privilege levels does not prevent information leakage via timing anomalies in interrupt handlers.
Black Hat Asia 2026 | Model Files → Memory Corruption → RCE: The Triple-Stage AI Attack Chain
Researchers Lu Lei and Zhou Jian demonstrate a triple-stage attack chain transforming model files into memory corruption and ultimately remote code execution within PyTorch. Their analysis reveals that enabling the 'weights-only' security flag in PyTorch's load function fails to prevent exploitation due to a logical flaw in the internal 'torch.jit.load' mechanism. This flaw allows malicious operators, such as file writing and reading functions, to execute arbitrary code despite safety configurations. The study highlights integer overflow vulnerabilities in TensorFlow and logical bugs in PyTorch that bypass established security assumptions, posing a critical risk to deployed AI applications.