I believe when you upload your code it will throw an error that your source and destination bucket are the same, which will prevent this type of looping.

This is not the case. I've written a Lambda which quite purposefully writes to a different path in the source bucket.

I have a Lambda for processing large files; the first phase is a single Lambda splitting the file into chunks which can be processed within the 60 second limit, and then the second phase is a bunch of the same Lambdas processing those chunks.

(Disclosure: I work with nicksergeant at Localytics.)

This wasn't the case when I was uploading mine, but maybe they've done some sort of smart checking on the execution role (it'd be pretty hard to detect just by looking at the code, I would think).