I noticed that Bayes is picking-up on very strong tokens from "eval" and "code" in headers like this:
X-PHP-Originating-Script: 1013:global.php(1938) : eval()'d code The "eval()'d code" part is in just over 2% of my spam, but it's never occurred in a single ham in my corpus. The spams seem to be coming from exploited web-servers, and I'm wondering if it might be a symptom of the exploit.