Can AI models exploit timing side channels? Devs clash over the idea
dcolascione · x · 2026-10-24
A debate over whether AI models could exploit timing side channels to infer their environment or leak secrets. dcolascione argues it's not absurd — cryptographers constantly worry about timing side channels — though he notes he's seen AI produce rigorous formal proofs (ASAN vs non-ASAN equivalence as a static memory checker) and calls this attempt "textbook" wrong.
Follow-up points: neural nets have a basically fixed amount of computation, models have no access to underlying floats and can't measure speed to infer the environment; plus inference GPUs are typically separated from the CPU running the agent harness, so an "AI agent" may physically span GPUs in different locations.