Introduction

Background

Malicious code, also referred to by common terms such as viruses, worms, and trojans, are a significant component of the scope of attacks that a modern IT organization must be prepared to defend against if they are operating with any Internet connectivity at all. The general term of malicious code, an umbrella term, is used to describe any code that performs unsolicited activity without the authorization of the user, and the more common and specific terms are often seen in technical write-ups of specific instances, or in the press due to their wide spread recognition.

Worms, a specific case of malicious code, are sometimes called blended threats and differ from traditional viruses primarily because they are able to propagate across a network with no intervention from the user. Whereas traditional viruses generally are defined as requiring some interaction with the user, possibly running a program, or perhaps clicking on a link to download some component via a Web site; worms have the capability to spread with no user interaction at all; most often this involves searching out hosts that are running vulnerable services and exploiting these vulnerabilities. Worms are especially dangerous because they require no action from a user; therefore if there is an outbreak, they are able to spread far quicker, and the spread is not confined to business hours in the affected region of the globe.

The discussion of worms and worm propagation strategies has been ongoing for a number of years, but it has been primarily in the last three years that the true significance of what a worm can do has been recognized in the more mainstream areas of network security and virus detection. The significance of worm propagation was truly noted in a wide spread fashion with the outbreak of Code Red in July of 2001. In addition to propagating, Code Red also performed a Denial of Service (DoS) attack on a single IP that, at the time, was registered to the White House. Since Code Red, the development of more complex infection strategies and payload components has proceeded to the point where the W32.SQLExp worm in January 2003 achieved worldwide propagation far faster than any human initiated intervention could have stopped the threat.

One outgrowth of this focus on worm activity is that a lot of research and analysis of vulnerabilities and worm propagation techniques has been published to aid security practitioners in evolving their policies and technologies. When examining this research, a couple of noteworthy things can be seen. Most of the research that has been pursued has looked at the spread of worms in the context of the global Internet rather than the rate that it will spread through an isolated private sector of the Internet, a corporate network for example. When looking at a global threat, and trying to estimate response time with given protection scenarios, this difference can be critical.

The importance of how a worm behaves in a limited, isolated, sector of the Internet gains importance when you consider that many organizations do not have enough publicly routable IP addresses, and therefore use private networks (eg. 10.X.X.X, 172.16.X.X, 192.168.X.X) behind network address translation devices for a majority of their Internet-connected computers. As these systems are not directly connected to the Internet; worm infestations must spread to these systems from already compromised internal hosts, and therefore the propagation behavior of worms on LANs can be a significant issue. The same issue can be seen even if an organization has public IP addresses, but employs strong firewall rules. Traditionally, many people felt that ensuring a firewall was in place and protecting hosts was enough, but as was shown by W32.SQLExp and Blaster, this is a flawed assumption.

Worm Propagation Factors

Numerous analysis documents and projects have focused on the propagation factors of various worms in an attempt to determine the most influential factors in the propagation rate, and to determine the strategies that are likely to be used by worm authors in the future. The paper by Tom Vogt entitled "Simulating and optimising worm propagation algorithms" is an attempt to do exactly that. Vogt builds a simulation network to test the impact of various strategies on the overall rate of propagation. Throughout these tests Vogt outlines some of the factors with the largest impact on the simulations of propagation.

  1. Address selection: The method of address selection is of tremendous importance in the overall rate of spread across the Internet. Different methods can include fully random, local preference random, and sequential scanning.
  2. Threading: A single thread of scanning results in a significantly slower rate of propagation than multiple threads, simply because the thread usually blocks until a scan or infection is complete before moving onto the next host.
  3. Pre-scanning: Doing pre-scanning to determine if a host is listening on a given port prior to sending data to the host is more efficient. This is only applicable for connection oriented services, or services that require a large amount of data to be sent in order to compromise.
  4. Method of scanning or infection: The use of efficient routines to minimize the wait time for infection or scan results. An example of this wait would be lowering the timeout variable if using TCP and the standard socket library.

In Vogt’s paper, he comments on the impact of these different factors in global scale propagation, but pays little attention to the impact that these factors will have in a more localized, isolated environment that is a component of a large global environment. Organizations with connectivity to the Internet should consider the impact that different strategies are likely to have on the local environment, as some strategies inappropriate for rapid global spread can have far greater significance when looked at from the perspective of a localized environment.

Why Analyze Old Worms?

Worms released since the first Code Red worm provide an excellent opportunity to examine different strategies and vulnerabilities that have been successfully leveraged to build a successful worm. While each of the worms employed different strategies resulting in significantly different propagation rates, they all achieved a substantial degree of success, often finding a small crack to penetrate an organization’s security posture. Analysis of these worms, why they were successful, and an examination of the protective strategies advocated in hindsight can allow organizations to better position and prioritize network security spending to protect against possible future worms.

As additional motivation for considering this topic, we just need to look at the "Simulating and optimizing worm propagation algorithms" paper. In the paper Vogt briefly discusses the impact that a truly malicious payload would have when combined with an effective and fast spreading worm. He models the effect that two types of malicious payloads would have, one being a randomly timed deletion of the hard drive, the other being a randomly timed DoS attack on a given Web site. In both cases the result would be catastrophic, and likely result in far greater amounts of lost productivity for any infected organization than anything previously seen. In his paper Vogt refutes the often cited, but ill-conceived wisdom that a worm that destroyed its host would not propagate successfully. This is not an attempt to glamorize or threaten IT organizations, but rather to highlight the first significant modeling of propagation with a randomly timed destructive routine. The most telling statement made by Vogt is:

    “At the end of this simulation run, there were 2,774 infected and 1,979 not infected systems left, of an initial 166,730. While this not quite "annihilation", it does mean that within two minutes, 161,977 hosts or about 97% of the vulnerable population were wiped out.”

The fact that a worm has not been written that successfully spreads and contains a highly destructive payload is fortunate, but currently there are vulnerabilities that could provide a successful exploit path for a worm such as this. In the Conclusions section, we will be providing an analysis of these vulnerabilities, a suggestion of the possible form that a worm might take for each vulnerability, and discussion about which of our recommended approaches are likely to pay dividends and control widespread infection.

The wide variety of services that the worms analyzed in this document target makes for a strong comparison of the effect of various protection strategies. As these worms all attack a widely deployed network service, but each has a different network footprint, we are able to compare the effectiveness of:

  • Network perimeter firewalling/router ACL application
  • Intrusion Prevention or Detection Systems (IDS/IPS)
  • Host based firewalls
  • Aggressive/fast patching strategies
  • Stronger network segmentation

Each of these solutions has been cited as a panacea when examining the impact of the worm in hindsight. In reality, each strategy has a limited applicability and a distinctive cost in terms of dollars and of service impact. Additionally, some of these strategies are completely unfeasible for certain vulnerabilities or environments (we see this primarily in the analysis of patching for W32.SQLExp).

Which Worms to Analyze

The choice of appropriate worms for this analysis was difficult, and in the end, we chose worms that affected Windows systems primarily because they were widespread, long-lived, and each had a distinctive difference in the attacked services network footprint. While the fact that Windows systems have been affected by the largest array of worms is obvious, the Threat Analyst Team does not believe this is due to any inherent superiority in other platforms, rather it is indicative of the density of similarly configured hosts.

Code Red I Family

The Code Red family of worms targeted a port hosting a widely deployed public service, HTTP. The Code Red worms leveraged a vulnerability in the Microsoft IIS HTTP implementation that allowed arbitrary code execution with the privileges that the IIS server is running under, usually LocalSystem. The use of a widely deployed service on a port that many firewalls and network perimeters leave open ensured that organizations that had not yet deployed a patch, and were unable or unwilling to block TCP port 80, were likely to be infected.

The vulnerability leveraged by Code Red and subsequent variants occurs in the processing of .ida files, used by Indexing Services. The overflow occurs in processing of a request, usually in the form of a URL encoded GET request.

The worm itself began spreading on July 12, 2001, and two and a half years later, the original and its variants remain a problem on the Internet today with infected systems continuing to attempt to propagate the worm. Other than systems that are badly administered, another common problem are hosts installed from original media, and infected before the patches can be downloaded and applied. This is a more significant problem for home users, but should not be ignored by administrators working rapidly to re-install a system that was providing IIS services.

Code Red I contained a flaw in its address selection, and a second version of Code Red I that fixed this particular flaw, dubbed Code Red version 2, was released shortly after. A third worm, Code Red II, substantially different and not based on the original worm, inherited the Code Red moniker, and will not be examined in this analysis. Instead we will be confining our analysis to Code Red versions 1 and 2.

One interesting fact about the Code Red I worms was that they existed solely in memory; no files were written to disk, and if a machine was rebooted, this cleared the infection from the system until the next exploitation. Once discovered, this memory-only existence allowed administrators to apply the patch, reboot the host, and no traces would remain of the compromise.

W32.SQLExp (also known as SQL Slammer or SQL Sapphire)

This worm, which was given multiple names by different organizations including W32.SQLExp, SQL Slammer, and SQL Sapphire, was notable in that it spread with a single UDP datagram. The use of UDP ensured that the worm could essentially use a fire and forget mechanism for propagation, and spread at a rate greater than any worm previously seen.

The worm spread to systems running the Microsoft SQL Server engine, and took advantage of a buffer overflow in the SQL Server Resolution Service, which operates on UDP port 1434. The overflow can be triggered by a single UDP datagram, meaning that the only way to mitigate against this is to drop the offending packet before it reaches the service.

By sending a single UDP datagram to randomly selected hosts, W32.SQLExp was not confined by having to wait for any particular blocking socket call to return to continue propagation, instead the sendto() function returns as soon as the data is written into the outbound queue. The use of UDP is beneficial when we consider the research by Vogt that showed a significant portion of the time spent during propagation in many worm designs is spent waiting for responses from vulnerable or non-existent hosts, and the use of UDP removes this bottleneck.

Various analyses of the global propagation rate, both theoretical and observed, indicated that a majority of the vulnerable systems had been infected within the first 15 minutes of worm propagation, a testament to the speed of the fire and forget method of exploitation.

W32.Blaster

W32.Blaster targeted Distributed Component Object Management (DCOM) over Remote Procedure Calls (RPC). Windows RPC is used to allow a remote system to use a service operating on a remote system without having that service be bound to a standard port, and without having to know the port in advance. The DCOM implementation allows software on remote systems to request components from a host, without those components running at the time. RPC is used by many system management tools in Windows networks, and is installed and operates by default on almost all Windows systems.

In August 2003, a vulnerability was disclosed that allowed stack memory to be overwritten by a malformed packet. The problem occurred in the processing of a network/file-system path for a DCOM request. The development of an exploit that was robust on different service packs and language versions took some time, but once an exploit was released that contained a so-called universal offset, the groundwork for a robust work was present.

The worm itself took this universal offset exploit and used other tools, including TFTP to transfer the worm binary onto the infected system. In comparison to the Code Red and W32.SQLExp worms, W32.Blaster was quite rudimentary, and it was surprising to some people that it was as effective as it turned out to be.

Network Configuration

In order to compare and contrast the effect these worms have had, and different paths through which they may have gained access to a network, we need to establish a "typical" network configuration consisting of the various systems that were most often cited as the access point for the worms. For the purposes of this paper, the network in question is logically segmented into a DMZ protected by router Access Control Lists, and an internal network that is behind a firewall that performs network and port address translation (NAT/PAT) in addition to proxying many protocols from inside the network to the Internet. The following diagram outlines the basic topology.

Figure 1. Network configuration border design
Figure 1. Network configuration border design

Email services are provided by Exchange servers running behind the firewall, and external email is relayed through a Unix system in the DMZ, thus ensuring that the Exchange systems are not exposed directly. The exchange systems are running IIS in addition to Exchange.

HTTP Services for the public and clients are deployed in the DMZ with application and database servers also in the DMZ. The DMZ is protected by a border router that allows only TCP 80 (HTTP), 443 (HTTPS), 25 (SMTP) through to select systems. Access to the DMZ from the corporate LAN is relatively permissive to allow management of data on the DMZ. But connections from the DMZ back inside the corporate LAN are confined to a very limited number of host/port pairs to ensure the integrity of the internal network.

VPN connections are provided for off-site employees, and terminate on the firewall. The VPN system is designed such that the client will ensure that all network connectivity must go through the VPN while the VPN is connected, this ensures that the system cannot be easily used to proxy compromised traffic from a remote system.

The corporate network contains 5000 desktop or workstation computers, 400 servers providing email, file and printer storage, and network management (Active Directory). This includes servers that are deployed for test deployment, development, and other purposes. The desktop systems are a relatively balanced mix of Windows 2000 and XP, the servers are all Windows 2000, Exchange 2000, and SQL Server 2000. Laptops and other mobile devices are placed directly on the protected LAN when the employees are in the office. This opens the possibility of an infected laptop provides an alternative path of compromise for worms.

The DMZ network contains 150 systems, primarily Windows 2000 Server, most with IIS installed and configured to allow remote administration or Web services. The router allows connections to a few of the HTTP servers only. A single SMTP host that includes attachment stripping is used to sanitize and forward to the internal corporate mail servers.

While it would be possible to add significant complexity to this network design, both in terms of services provided and security precautions taken, the design is sufficient to model the risks that the analyzed worms present.

Analysis

Code Red v1 and v2

The Code Red worm, both version 1 and version 2, employed a random address selection routine, although weaknesses in the initial seed of the worm meant that version 1 of the Code Red worm attempted to spread to the same list of IP addresses on each run of the worm. Version 2 fixed this initial seed weakness, and consequently saw far greater spread across the Internet, with hosts that were untouched by the first version becoming victims of the second one.

One observation of the rate of SYN probes sent out by a Code Red version 2 infected server was eleven packets per second [ref 1]. We can assume that this rate of propagation was also noted in Code Red version 1, but because of the weakness of initial address selection, it may have been possible for a system to not be scanned. For the purposes of our calculation, we will assume a random address selection with 11 probes per second each to a unique address.

In our example network, Internet access to port 80 is blocked to all systems in the corporate network, and is allowed only to systems providing Web services that are located in the DMZ. Of the servers providing public Web services, if we assume that most had the patch successfully deployed, and therefore, just a single host was vulnerable and listening on this port, we can discover the risk from a minimal infestation. Due to the cost of patch deployment, only systems that were deemed “at risk” of exploitation (meaning those that were exposed remotely) had the patch deployed. There were a number of other servers in the DMZ, including SQL servers, Domain Controllers, and monitoring systems that were running vulnerable IIS, but these were not accessible from the Internet due to perimeter routing. Even though these servers are not accessible from the Internet, they are still accessible once the worm has penetrated the perimeter.

Once this publicly exposed server is attacked, and compromised, the scanning routine starts. The infected machine will randomly select IP addresses to attempt to connect to, at an average rate of 11 packets per second. The address range for this random selection is the complete Internet, and the chance that this single infected host will infect other hosts on the same network is relatively remote, around 1 in 4.2 billion. It would take over 12 years for a single infected host to traverse the address range of the Internet. Assuming that no other hosts are vulnerable and remotely accessible, the speed of an outbreak on the local network will be quite slow.

For random scanning propagation, we can use the calculation of:

            AddressRange
    -----------------------------
    2 VulnAddress PropagationRate
    

Where:

  • AddressRange is equal to the complete Address Space that is traversed in the scanning routine
  • VulnAddress is equal to the number of hosts in the network that are at risk
  • PropagationRate is the number of vulnerable hosts per second that are targeted by an infected host

This calculation was derived from the statistical fact that assuming a randomly dispersed brute force guessing routine, it takes an average of n/2 attempts to determine a given value from a pool of n size. If you divide the size the pool by the number of possible correct values, you will determine the number of guesses, on average, that it will take to guess a single correct value. Once you take into account the number of connection attempts it can perform per second, on average, you can determine the average time it will take.

For Code Red v2, the calculation is:

     253*255*255*255
    -----------------
        2*100*11
    

This type of random address selection, although effective at spreading a global worm widely, and utilizing the exponential growth of numbers of infected hosts to fully compromise the Internet, is not terribly effective at spreading inside a protected network once it has penetrated the firewall on a single system. The fact that every address in the IPv4 address space has an equal chance of infection means that there is a relatively small chance of infection of any specific address or address ranges. The Code Red worm is a perfect example of a worm that was designed for global reach and global propagation, rather than designed to cause problems on a local network once it has penetrated the perimeter.

For organizations that had deployed patches to as many of their systems as they could, and had reasonably configured firewalls / perimeter filtering, Code Red, and worms that did not have an optimized propagation algorithm would be a minor threat to their operations, and the outgoing connections to the Internet from the infected system would likely identify this system before a significant number of other internal hosts were affected.

In the case of Code Red, a worm that targets a vulnerability in a widely deployed service that must remain open, border filtering to limit the infection is not always feasible. Despite the potentially slower infection rate due to random scanning, hosts on the interior of the network that did not have the patch in place were at risk, and should an infection gain a toehold, the consequences would have been significant. This vulnerability certainly proved that patch application for all systems, not just those exposed to the uncontrolled networks, needs to be performed. Deployment of a personal firewall might have been a reasonable option to limit the exposure of systems that have IIS installed by default, but do not use the standard HTTP port for any service.

W32.SQLExp

The W32.SQLExp worm was a generation ahead of the Code Red worm mainly due to the speed with which it spread. As was outlined earlier, many estimates show this worm compromising almost all the vulnerable, addressable hosts on the Internet in less than 20 minutes. This is significantly quicker than the speed that Code Red spread at.

The address selection employed by W32.SQLExp was in the same vein as that used by Code Red versions 1 and 2. It was a random address selection. As identified in the “Simulating and Optimising Worm Propagation Algorithms” paper, this is not the most efficient approach, but it was effective.

W32.SQLExp was optimized extremely well, but did contain weaknesses in the address selection that meant that any given infected host would not be able to propagate to the complete Internet.

As with Code Red, W32.SQLExp was optimized to spread globally in a rapid fashion, with a single infected host behind the protected network perimeter, each selection of a new host would have a 1/4.2 billion chance of hitting another vulnerable host on the network. Because the worm spreads at wire-speed, instead of taking 12 years to completely traverse the Internet address range as Code Red was subject to, W32.SQLExp can traverse the same address range in 48 hours (assuming no duplication of addresses, in reality this is very unlikely, but it highlights the difference in propagation speed on a single host).

Worm size = 404 bytes UDP + 30 bytes Ethernet framing = 434 bytes.

On a 100Mbit switched full-duplex network, at 80% utilization (a conservative estimate) 23,041 W32.SQLExp UDP packets can be sent per second. If this were a 1000Mbit connection, the amount of time to spread would be correspondingly lower, likely around 5 hours to traverse 4.2 billion addresses.

Using the same equation that was used for Code Red, and assuming that 150 hosts are vulnerable (possibly running Visio or other product that shipped MSDE, or were deemed lower on the patching priority list) in the network.

     255*255*255*255
    -----------------
       2*150*23,041
    

There would be an average of 612 seconds, between W32.SQLExp gaining access to the protected network via a host 100Mbit network, and the secondary infection. Once another host was infected, this interval would be halved, and it quickly becomes apparent that complete compromise would occur in under an hour. The resulting traffic from the scanning would be significant, and would likely saturate the LAN and the outgoing connection causing tremendous disruption.

It is impossible to accurately gauge the range of addresses that W32.SQLExp would arrive at for propagation, due to weaknesses in the address generation routine identified in the analysis. The point of this exercise is to demonstrate that a single host that is connected to the internal LAN once it has been infected W32.SQLExp was a far greater risk than a single host compromised by Code Red.

The other complicating factor with W32.SQLExp was that there were many more applications that were at risk due to the distribution of the Microsoft Desktop Engine, and that many applications that distributed MSDE did not have updates available which made successful patching an incredibly difficult, and sometimes futile exercise. One point that should be considered is that on most installations of MSSQL, the UDP 1434 port was not required for proper functionality, and it certainly was not needed on most MSDE installations. The fact that this port was not required would seem to lend weight to a firewall solution.

In our sample network, appropriate border filtering, denying incoming traffic on this unneeded port, ensured that the DMZ was not infected from the Internet. A mobile laptop was the greatest risk, and once it had been moved, logically, behind the filtered perimeter, it was able to infect a significant number of other systems. Additionally, because administration of the hosts in the DMZ was important, traffic from the corporate network, thought safe, to the DMZ was relatively unfiltered and therefore the SQL servers deployed in the DMZ could also be infected from a single laptop or VPN user.

Personal firewalls for these mobile computers would have reduced the risk significantly, but due to the speed that this worm spreads, it takes just a single infected host to compromise all the relatively unprotected systems behind the perimeter. Another option to consider would have been personal firewalls deployed across the enterprise. While this option would certainly control the spread of a worm like W32.SQLExp (on most systems that were vulnerable UDP/1434 was not required to be open), the difficulty with managing a personal firewall rollout of this size would be significant.

One approach that would pay dividends is to treat mobile computers, whether through the VPN or laptops as untrusted and possibly hostile. Adding another network segment, a mobile computer zone, would mean that filters could be put in place to control the access that these computers have to the corporate network. Appropriate filtering rules (allowing only connections to necessary servers for example) would mean that those critical servers could be patched first in a patch management scheme, ensuring that should a worm begin to spread, it can be contained to a specific network segment.

W32.Blaster

The Blaster worm was incredibly slow spreading globally, and some commentators questioned whether it was a significant worm because of that. The slow speed was especially evident when comparing Blaster to the speed of W32.SQLExp’s propagation. One surprising element of this slower spread was the fact that many organizations found their Internal networks to be affected far more negatively than was seen for earlier worms.

Blaster propagated in a fashion that tended to favor addresses numerically close to the infected host. Blaster generates a random number modulo 20, and if the result is greater than or equal to 12, it bases its scanning off of the local address, if it is less than 12, it picks a random network to begin scanning. Once the scanning routine has chosen a starting point, it sequentially scans, incrementing the IP address by one, and maintains this pattern until it is restarted, whereupon it once again produces a random number modulo 20. The result is that a host, until it is restarted, will scan sequentially from a given starting address with a 40% chance that it will pick an address related to the current IP.

The “Simulating and Optimising Worm Propagation Algorithms” paper notes that sequential scanning is incredibly inefficient at spreading a worm, and in this case, that assumption is born out as true. There is an interesting correlation though, and that is that because of the preference for local addresses, Blaster was far more effective at compromising every vulnerable host in a local network.

This is particularly notable when you consider that in this day and age of network security products, including antivirus, firewalls, change control policies, patch management policies, IDS/IPS systems, most organizations are not wide open to threats. Rather, there are two types of vulnerable hosts that are at risk of infection, those that are exposed to the Internet and remain unpatched, and those that are not patched or well hardened but are well protected from the external environment by firewalls. In many situations, the real risk to most organizational networks these days is the interaction of these two classes of vulnerable hosts.

When we examine our example network, a concerted patching effort was implemented due to the wide spread exposure of TCP/135, especially for desktop systems, but mysterious, unidentified failures in patching left just short of 25 percent of the systems vulnerable to attack [ref 2]. Personal firewalls were also strongly advised, especially those high risk mobile PCs.

One complication is that the RPC port is used for managing systems in a Windows environment. Therefore deploying a personal firewall on all the systems will likely require a number of systems with exceptions, and as not every application fully documents what is required to operate correctly, therefore troubleshooting some configuration problems may become far more complex.

As with W32.SQLExp, the best approach is likely to try and segment the network with mobile hosts segmented from the other internal network. One concern with this approach is that if one of the servers that is running Exchange had a silent failure with the patch application, then 135 would be exposed and the worm can hop past the firewall segmenting the mobile network.

Conclusions

Summary

Worms are making advancements that optimize the rates that they spread at, and increasingly, a successful worm can be measured in two fashions. Worms that are highly efficient at global Internet spread are typically much slower at fully compromising a local network with a single infected host that is behind the perimeter filtering. Therefore, the fastest spreading Internet worms, Code Red and SQL Slammer for example, are less efficient at fully compromising a LAN than a worm that has a preference for local addresses, evidenced by the W32.Blaster worm. It is possible for a worm to pursue a hybrid model, where a network is chosen randomly, and then that network is sequentially scanned.

There are some strategies that, once employed, can narrow the gap in efficiency in both cases, UDP based worms, or worms that do not rely on standard TCP stack communication will be able to spread faster, because they do not have to wait for a connection information, such as the SYN+ACK or RST of TCP, to occur. This effectively allows the system to saturate the outgoing link, and not become slowed down by some of the strategies that have been suggested and implemented (such as Tarpitting). Ultimately though, schemes that scan randomly have far greater spread across the Internet, while sequential scans will have far greater impact on the local area network.

Recommendations and Looking Forward

Organizations must begin to address the weaknesses that are inherent in the topology that has been commonly deployed in the past. The notion that there is a safe local network, and a hostile external network is a misnomer. As devices become more network aware and mobile, this differentiation between internal and external machines, trusted and untrusted will continue to erode. As additional sites and organizations connect via VPN connections, demands are made to provide more access to services, we can expect overall system security will decrease.

Application proxying, often used these days for HTTP and SMTP services can certainly help. And we can expect more proxying and filtering technologies are deployed to support new network services, such as Instant Messaging and NetPhone, Video Conferencing applications.

While proxying technologies will advance, one of the more profound changes will be in embedding filtering and firewall services in the network fabric itself. As the move to switched and VLANed networks helped to increase the efficiency of many organizations networks, the move toward this emerging technology that can control access to the network based on policy compliance will have an effect on helping to control outbreaks of worms on networks with solid perimeter protection. In the meantime increasing the segmentation and filtering between network segments is likely to limit the actual worm exposure of the whole network.

Workstation Service

On November 11, 2003, Microsoft released Security Bulletin MS03-049, which outlined a vulnerability discovered in the Workstation service. This vulnerability is the fifth remotely exploitable vulnerability found that is accessible through the RPC Interface in Windows systems. The workstation service is used to authenticate and request services from other Windows hosts, and therefore, in a corporate environment it is impossible to fully block or turn off this service.

The essential nature of this service means that sites that utilize VPNs and laptops will be unwilling to filter the port, meaning that machines that move between the protected LAN and other networks will be at risk.

Further analysis by the discoverers of this vulnerability, eEye, indicates that this is a stack based overflow, making exploitation significantly easier for a number of exploit authors, and they outline some details of the vulnerability that will likely speed the creation of public exploits.

In recent history, most worms are built around known, publicly released exploitation mechanisms, and in general, the release of public exploits, and the reliability of these public exploits, is correlated with the chances that a worm will be written to exploit the vulnerability. Despite all of these factors supporting the creation of a successful worm, at the time of this writing an exploit that is reliable across multiple service packs and versions of Windows has not been written. In addition, the only Windows 2000 exploit available relies on the file system being a FAT32 rather than NTFS.

Messenger

On October 15, 2003, Microsoft released Security Bulletin MS03-043, which outlined a vulnerability discovered in the MS Messenger Sub-System. This system is available on every deployed version of Windows, and is listening by default. The system is used to send pop-up notifications to users about events that occur on the network, and is utilized in print functions, network administrator information broadcasts, and other functions.

The vulnerability, upon further investigation, appears to be a heap overflow, and if exploitation occurs, it will be more difficult to do reliably than the stack based overflows that resulted in all the worms analyzed in this document. The fact that this vulnerability has been released since October 15, 2003, and a successful exploit has not yet surfaced publicly indicates that this vulnerability may not be the fruitful opportunity for worm development that it was originally thought to be.

This vulnerability is present over UDP, and is accessible through a number of ports, most commonly UDP/135 and an ephemeral port, usually UDP/1026.

A worm exploiting this service would be incredibly fast spreading owing to the use of UDP as the network protocol, meaning that it could utilize the strategy employed by W32.SQLExp to fire and forget exploitation packets. Should an exploit be discovered that utilizes this strategy, it is likely that this will be an incredibly fast spreading worm, and the Threat Analyst Team believes that there is a reasonable chance that exploitation will occur over 1026 as well as, or rather than, 135, simply because activity over 135 means that this port has far greater scrutiny.