How Vulnerable are Unprotected Servers in the Cloud? Part I

Written by Grace Zeng. Posted in Blog Post

By Grace Zeng and David Coffey

The Internet is a playground for opportunistic attackers. Right now, there are thousands of malware threats circulating around the Internet. Most computers today are protected by firewalls, IDS/IPS and anti-virus (AV) tools. But what happens when they do not have any protection? Previous experiments on “Time-to-Live-on-the-Network” and “Survival Time” of Windows machines were conducted quite a few years ago with test machines running old Windows operating systems. The “Four-Minute Windows Survival Time” claim in 2008 was especially criticized for using a Windows XP RTM or SP1 version in the test.

Since the time of these initial time-to-live studies, the Internet threat environment has become deadlier. Meanwhile, the Windows operating systems have become more secure. Because the state-of-the-art changes so quickly, we wanted to know how well an unprotected machine with a current operating system does in today’s threat environment. Left to its own devices, how soon will it be probed and attacked? We are particularly interested in testing unprotected machines hosted in the cloud because enterprises are increasingly turning to the cloud for various business purposes.

Experiment design

We ran our experiment with 15 machines in Amazon’s Elastic Computer Cloud (EC2) environment with two configuration profiles: “wide-open” and “out-of-the-box”. In the wide-open scenario, a machine opens all ports and emulates all possible services. This way the machine can attract as many malicious attempts as possible. In the out-of-the-box scenario, a machine runs only with default open ports and services. This scenario gives us a baseline of how many malicious attempts an unprotected machine might encounter.

Windows is by far the most popular operating system on the Internet. Its Server versions are generally exposed to more risks than Home/Professional versions. Our tests were carried out on the latest Windows Server 2008 R1 SP2 and R2 SP1. We disabled all firewall and anti-virus programs and configured the security policies so that Amazon would allow all incoming connections (TCP, UDP, ICMP) to those machines. We used Wireshark to capture packet-level traffic in real time.

To create the wide-open scenario, we installed a low-interactive honeypot named HoneyBot and disabled/changed several services to avoid interference. After the configuration was complete, we took a snapshot of the instance and created an AMI (Amazon Machine Image) for later use. We launched ten instances on EC2 using the same AMI and made sure that they were hosted in different geographical zones and were allocated different IP addresses.

For the out-of-the-box scenario, we made a clean install of Windows Server 2008 and didn’t install any programs other than Wireshark. By default, only ports 135 (RPC), 139 (NetBIOS), 445 (SMB) and 3389 (RDP) were open. We ran five such instances on EC2.

Scan, probe and exploit elapsed times

Malware infections follow a predictable pattern. Using a port scan, an attacker tests whether a port on a target machine is open. If so, a vulnerability probe gathers more information about a listening service, such as the version of the service – to identify vulnerabilities; and an exploit delivers malicious payloads to compromise the machine.

In  the wide-open scenario, after launching, on average it took about 23.4 minutes to see the first port scan, and 56.4 minutes to see the first vulnerability probe (the number for each server is shown in Figure 1).  Probes hit well-known ports such as 22 (SSH), 23 (Telnet), 25 (SMTP), 80 (HTTP), 445 (SMB), 1080 (SOCKS Proxy), 1433 (Microsoft SQL Server) and 3389 (RDP). With respect to exploit times, we observed that almost all first exploits were made within 24 hours, with the average time being 18.6 hours (Figure 2). We captured exploits on port 445 (SMB), 1434 (Microsoft SQL Monitor), 2967 (Symantec AV) and 12147 (Symantec Alert Management System 2). Almost all exploits during our month-long experiment were known threats. For example, the attack targeting port 12174 exploits a remote-code-execution vulnerability which was disclosed in 2009.

Figure 1. Scan and Probe Times for Wide-Open Servers (in minutes)

Figure 2. Exploit Times for Wide-Open Servers (in hours)

In the out-of-the-box scenario, it took an average of 13 minutes for the first port scan to arrive (Figure 3). Port scans hit ports such as 8080 (HTTP) and 1433 (MS SQL Server). The first vulnerability probe arrived within 3 hours on average (Figure 3); all probes were login attempts to Samba share (445) or via RDP (3389). We monitored the servers for a few weeks but didn’t see any exploits due to the limited number of open ports.

Figure 3. Scan and Probe Times for Out-of-the-Box Servers (in minutes)

Wrap up

Back to the question we asked in the beginning: On today’s Internet, how long does it take for an unprotected machine in the cloud to be probed and attacked? The short answer is: not very long.  What do we learn from the experiment? What can you do to beef up the defense of your machines? Go and check out Part II.

Andrew Jaquith and Richard S. Westmoreland contributed to this post.

Tags: , , ,

Trackback from your site.

Grace Zeng

Yuanyuan Grace Zeng is a research analyst with Perimeter E-Security. She recently received her Ph.D. in Computer Science and Engineering from the University of Michigan, Ann Arbor. Her specialty is network security with focuses on malware modeling/analysis, attack detection and mitigation as well as large-scale data analytics. Her work on botnet detection has been published in several premier security conference proceedings.

Leave a comment