The time-warp attack, and the timestamp rules that bound it

The first assumption is a majority of hash rate. Without it, the miner does not produce enough of the blocks in a retarget period to control the timestamps the adjustment reads. This is the same assumption the 51% attack page states, and it is the binding constraint on the strategy.

The second assumption is that the timestamp rules do not close the gap. The median-time-past rule bounds how far back a miner can set a timestamp, and the two-hour future rule bounds how far forward. The strategy can only move a timestamp within that window, so the difficulty reduction it can achieve per period is limited. A change to the timestamp rules would change the attack's ceiling, which is why the rules are part of the security argument.

The third assumption is that the manipulation is not detected. A chain whose timestamps are visibly compressed is a signal that something is wrong, and the response available to the rest of the network is to reject the chain or to coordinate against the miner. The model treats the miner as unopposed, which is a simplification.

The defence is the timestamp rules themselves, and the cost of a majority. The median-time-past rule prevents a miner from setting a timestamp arbitrarily far in the past, and the two-hour future rule prevents one arbitrarily far in the future. Together they bound the manipulation to a window that limits how much difficulty a miner can shed per period. The difficulty adjustment's own clamp — a retarget may at most quadruple difficulty or reduce it to a quarter — further limits how fast the change can compound.

The practical reading is that the time-warp attack is a real result about the looseness of the timestamp rules, and a strategy that requires a majority to mount. The difficulty adjustment page sets out the retarget rule in full, and the security budget page explains what a majority of hash rate costs to acquire.