You'd have to scan for include($tmp) and require(HTTP_REFERER) and *_once as well. If an exploit script can write to a temporary file, it could just include that later. Basically a two-step eval.
And it's even possible to hide remote code with workarounds like:
include("data:text/plain;base64,$_GET[code]");
Also, if your webserver has already been compromised you will not always see unencoded evil. Often the exploit shell is gzip-encoded. Think of include("zlib:script2.png.gz"); No eval here, still same effect.
