I'm not saying this is the answer, or even safe depending on your data, but I had a similar situation solved by it. I was using a DB_File to YAML to DB_File script that was running perfectly for months until is just broke one day. Running this on the plain text data fixed it.

$text =~ s/[^[:print:]]//g;

(Probably had inserted a bunch of invisible garbage from users cutting and pasting out of different apps.)