WiP: DROW: Training-Free Load Speculative Execution Attacks on Apple Silicon

Speculative execution attacks continue to undermine modern processor security by leaking sensitive data across architectural boundaries. While prior Spectre-type attacks often rely on explicitly training hardware components into exploitable microarchitectural states, this dependency on training limits practicality, requiring deep microarchitectural knowledge and careful gadget preparation. In contrast, this paper presents Spectre Blind-Bypassing (Spectre-BB), a new training-free Spectre primitive induced by a stateless load speculation feature, blind bypassing, found in Apple silicon. We systematically analyze data-flow speculation and uncover that blind bypassing violates memory dependency checks without requiring mistraining or architectural state manipulation. Building on this primitive, we develop \name{}, a new class of real-world, training-free speculative attacks that hijack both data and control flow. Our evaluation across different dependency patterns shows that Specter-BB offers significantly higher attack bandwidth. We demonstrate four practical exploits, including cross-page website data leakage, kernel and library gadget discovery, breaking ASLR/KASLR and ARM Pointer Authentication circumvention.