This is regression in introduced in #140. When a string in the metadata section contains invalid UTF-8 characters the behavior Python 2 is leave the string exactly as it appears in YARA, in Python 3 however the invalid characters are removed because Python 3 strings are not handled as bytes like in Python 2, they most have a valid encoding. PR #140 was an attempt to homogenize the behavior in both versions of Python, but it introduced this other issue.