If a program sent out a malformed packet, this would indicate an error in the sender's program. Let's assume that a receiving program reads the malformed packet and is able to process it. It results in something favorable (no clue how this would be determined, but let's go with it) and so the receiving program repeats whatever it originally did to get this malformed packet sent to it, so that it gets another malformed packet. Multiply this by thousands of other programs doing similar things to "hack" other programs into sending malformed packets to perform useful functions, and maybe you have something happening here.

The above process basically describes a genetic algorithm, which I don't think the Internet is. It also assumes that most programs are flexible enough to produce bugs like this - creating malformed packets. Does this mean we should stop unit testing code to allow more freedom from the programs? Plus, the above also assumes there is some way of ranking an outcome from a malformed packet as "favorable". Maybe that would be humans?