How does a NOP sled work?

Some attacks consist of making the program jump to a specific address and continue running from there. The injected code has to be loaded previously somehow in that exact location.

Stack randomization and other runtime differences may make the address where the program will jump impossible to predict, so the attacker places a NOP sled in a big range of memory. If the program jumps to anywhere into the sled, it will run all the remaining NOPs, doing nothing, and then will run the payload code, just next to the sled.

The reason the attacker uses the NOP sled is to make the target address bigger: the code can jump anywhere in the sled, instead of exactly at the beginning of the injected code.

A 128-byte NOP sled is just a group of NOP intructions 128 bytes wide.

NOTE #1: NOP (No-OPeration) is an instruction available in most (all?) architectures that does nothing, other than occupying memory and some runtime.

NOTE #2: in architectures with variable length instructions, a NOP instruction is usually just one byte in length, so it can be used as a convenient instruction padding. Unfortunately, that also makes it easy to do a NOP sled.

To add to rodrigo's explanation - Even with a NOP sled, the approximate location of the buffer in memory must be predicted in advance. One technique for approximating the memory location of is to use nearby stack location as a frame of reference. By subtracting an offset from this location, the relative address of any variable can be obtained.

SIDE NOTE: on x86 architecture the NOP instruction is equivalent to the hex byte 0x90 therefore a completed exploit buffer could look something like this:

| NOP sled | Shellcode | Repeated return address|

Seeing as if the EIP register points to any address found in the NOP sled, it would increment while executing each NOP instruction, one at a time, untill it finally reaches the shellcode