Cognitive overload as an attack vector

The human/machine parallel has a security consequence. Researchers have shown that,
just as humans make errors under overload, LLMs can be deliberately pushed into
cognitive overload by crafted prompts — and that this state can be exploited to
bypass safety mechanisms, with very high reported success rates against leading
models. This is a sobering demonstration that the cognitive parallel is not merely
analogical; it connects directly to the prompt-injection material in EB-4.