For example, imagine a bot programmed by some 5-10 person team at a hedge fund. They find that running sentiment analysis on twitter and news comments can accurately predict whether a given security will rise or fall in response to a fed press release. The profits are good, so the team manager moves a couple million into the algorithm. One night after work, a team member tells his programmer friend about the algorithm, and mentions some of the most profitable stocks. This friend goes home and programs another bot to corrupt the sentiment analysis dataset, by posting fake comments with properly tuned sentiment. The hedge fund bot reacts as expected, and now the friend has the power to manipulate the bot. He has outsmarted the bot and can take advantage of the high volume trading.
That might be a bit of a contrived example, but corruption in machine learning data is a very real problem. People are just starting to study it. [0]
[0] https://www.usenix.org/system/files/conference/usenixsecurit...