A new, stealthier version of a previously known Russian Trojan horse program called Gozi has been circulating on the Internet since April 17 and has already stolen personal data from more than 2,000 home users worldwide.

The compromised information includes bank and credit card account numbers (including card verification value codes), Social Security numbers and online payment account numbers as well as usernames and passwords. As with its predecessor, the new version of Gozi is programmed to steal information from encrypted Secure Sockets Layer (SSL) streams and send the stolen information to a server in Russia.

The variant was discovered by Don Jackson, a security researcher at Atlanta-based SecureWorks Inc. who also discovered the original Gozi Trojan horse back in January. One of improvements is its use of a new and hitherto unseen "packer" utility that encrypts, mangles, compresses and even deletes portions of the Trojan horse code to evade detection by standard, signature-based antivirus tools. The original Gozi, in contrast, used a fairly commonly known packing utility called Upack, which made it slightly easier to detect than the latest version.

This version of Gozi also has a new keystroke-logging capability for stealing data, in addition to its ability to steal data from SSL streams. According to Jackson, the keystroke logger appears to be activated when the user of an infected computer visits a banking Web site or initiates an SSL session. It is still unclear how exactly the keystroke logger knows to turn itself on and capture information.

Apart from those two differences, the variant is identical to Gozi, Jackson said. The Trojan horse takes advantage of a previously fixed vulnerability in the iFrame tags of Microsoft Corp.'s Internet Explorer to infect systems. Users typically appear to be infected when visiting certain hosted Web sites, community forums, social networking sites and those belonging to small businesses.

The original Trojan horse stole more than 10,000 records containing confidential information belonging to about 5,200 home users, companies, government agencies and law enforcement organizations before being detected. The server to which the data was being sent to had a very professional-looking front end that allowed users to log into individual accounts, view indexed data and get results from queries based on certain fields such as URL and form parameters.