Mithril: An Experiment in Adaptive Security
NCSA Security Research
This is detailed account of our experiences and accomplishments in our attempt to deploy and manage an adaptive security system. Our system consists of an intrusion detection system (IDS), an attack correlation layer and a response layer which reacts to specific attacks as well as enforcing an adaptive system policy. We studied various implementations and found that the open source Prelude Hybrid IDS would best fit our needs. Prelude uses both network and host-based, heterogeneous distributed sensors to generate alerts. Its reporting protocol is Intrusion Detection Message Exchange Format (IDMEF). It currently uses Snort as its primary network sensor and its own Prelude LML (Log Monitoring Lackey) as its primary host-based sensor. It supports a variety of other sensor types by linking them into the Libprelude communication layer. Essentially the library enables a sensor to convert its native alert format to IDMEF and allows a secure communication path to a Prelude Manager. Once an alert is generated by a sensor, it is sent to a Prelude Manager where it is loaded into the Prelude database. The front-end to this database is the web-based management console Prewikka. We modified Prewikka to incorporate and allow the enforcement of our system policy settings. The Prelude team provided a mechanism for correlating alerts using the standard Simple Event Correlator (SEC). Our response mechanisms for enforcement are script-based. We created Perl scripts for reacting to specific attacks and Cfengine for system-wide modifications. Here is a good general diagram of the overall framework for Mithril.
Prelude Hybrid IDS
The Prelude Hybrid IDS was chosen because it offered the most flexibility due to being both open source and compliant with a variety of existing sensor models. This allowed us to gather a variety of generated alerts from various sources and analyzing them in a centralized system. Not being committed to a specific IDS type, we effectively implemented the hybrid IDS as our general system for gathering data and correlating events. Here is a non-exhaustive list of the strengths. Since it is a hybrid system it is flexible and highly adaptive. Being open source, the code can be modified easily to fit our needs. Prelude is generally cross-platform except for Win32, which is in progress. It employs flexible IDS rules so that it ‘understands’ and readily incorporates third party rules. It is a hidden IDS in that it is not attached to an IP. It uses the standard IDMEF protocol for alert specifications. And finally, it supports a large number of existent IDS’s including Snort, Honeyd, Nessus VS, Samhain and LibSafe. When we first installed and configured Prelude, we used version .8. We later upgraded to version .9 as significant improvements and capabilities were added. Minimally, version .8 uses the communication library Libprelude, the Prelude database, Prelude Manager, PreludeLML, Prelude NIDS and the web-based console Piwi. Version .9 uses the communication library Libprelude, the Prelude and Prewikka databases, Prelude Manager, PreludeLML, Snort and the web-based console Prewikka. Other types of sensors are available for both versions, but we decided that this minimal configuration would serve our needs without adding further complications. We used the available Debian packages since our test platform is Debian GNU/Linux. Here is an overview of the basic Prelude components.
Libprelude is the foundation of Prelude IDS system. It is a C library that enables all of the Prelude components to communicate via SSL connections using the IDMEF protocol. These connections are created after the installation of a sensor using a utility called ‘prelude-adduser’. In version .8 the utility is called ‘sensor-adduser’, but the functionality is identical. Generally, there is a handshaking password exchange between the sensor and the manager. Specific details will be covered when we discuss the installation of a sensor. Libprelude also provides the ability for third-party software to link with it making them ‘Prelude Aware’. That is, they can communicate with Prelude components using IDMEF. This is typically used when building a sensor which is hooked into the Prelude IDS. A sensor can be built using C, Perl or Python. We will look at the specific details for coding a sensor in the ‘Installing and Building Sensors’ section. Basically, the required steps include initializing Libprelude, creating a Prelude ‘client’ which enables interaction with the library and finally destroying the client. Here is a diagram of the general layout of the library culled from Prelude documentation.
The Prelude database is the repository for all of the data gathered within the Prelude IDS framework. Primarily it contains all of the alert information including timestamps, source/target IP, users, severity, etc. The database also includes Prelude user settings and agent data such as heartbeats, configurations and locations. The Prewikka database added with version .9 provides tables for specific needs of Prewikka. Database reporting support includes MySQL, PostgreSQL and Oracle. We used MySQL. The database tables for both versions can be found here.
The Prelude Manager is the component which recieves all of the alerts from the sensors. It provides centralized management for a highly distributed system. It also enables scalability and survivablity by being distributable itself. That is, many Prelude Managers can coexist on a single system sending data to a centralized database. This allows sensors to report to Managers which in turn report to other Managers by forwarding the alert data. This partitioned architecture maximizes coverage with centralized analysis while preventing entire system failure due to a successful compromise. There is also a built-in capability for communictaion with counter-measure agents. Here is the general architecture of the Prelude Manager culled from Prelude documentation.
Prelude LML (Log Monitoring Lackey) is the log file analyzer which acts as a host-based sensor. It does this by watching specified log files, typically syslog, auth.log and messages, and uses PCRE (Perl Compatible Regular Expressions) rules to match lines in the file. Once there is a match, an alert is created and sent to its Prelude Manager via SSL tunnels. It normalizes the data preventing multiple entries for a single alert. Prior to transmission, the alert data is transformed into IDMEF in Libprelude. More information is available on Prelude-LML: Overview/PCRE.
In version .8, the primary network-based sensor was Prelude NIDS. But since by design it implemented Snort rules, it was deprecated for version 9 and Snort itself is used. Because of this, we needed to build a compatible version of Snort (2.4.5) linking it with Libprelude. We will look at this specifically in our discussion of the Snort-Prelude sensor. The NIDS sensor is responsible for real-time network packet sniffing and analysis. To capture network packets, it uses a modified version of Pcap. The sensor decodes the packet headers and stores them internally. Then it performs various validity tests which detect any malformed data. More sanity checks are done on the IP and TCP flags. If an anomaly is detected, the packet is rejected. After these tests come IP defragmentation and TCP reassembly operations. Finally packet data is analyzed by Snort using signature, protocol and anomaly based inspection methods. In the earlier version, Prelude NIDS analyzed the packets using various plug-ins. With version .9, Snort is used exclusively.
Finally, let’s take a look at the web-based console. In version .8, Prelude used a Perl implementation called Piwi. In version .9, a Python interface is called Prewikka is used. Basically, the views allow the user to monitor alerts as they are received, filter those alerts so they can be aggregated and view appropriate graphics for analysis. We needed to make modifications with each version so that we could correlate alerts and activate appropriate attack responses and enforce system policy changes. Her are views of each version. The first is the alerts view for Piwi and the second is the alerts view for Prewikka.
Alerts and IDMEF
Obviously, alerts provide the groundwork for any effective IDS and ultimately determine the reliability of a reactive system. At the base level, we need to determine whether we are getting the alerts that we need. The Prelude sensor rules are very comprehensive. In fact, at times we needed to modify configuration files so that we would not generate alerts that were generally meaningless to us. One specific instance was a Snort alert called ’BAD-TRAFFIC IP Proto 103 PIM’ which generated an excessive number of alerts. The alerts were being triggered by a harmless routine anomaly caused by Cisco routers. We filtered out the alert by commenting out the appropriate line in the rules file ‘bad-traffic.rules’ and restarted Snort loading the modified rules. We did this for a number of extraneous alerts as well. Also, the IDS is only as effective as its rules, so updating the rules on a regular basis strengthens the detection process. NIDS rules updates are available for Snort on its website and updates for LML are always up on the Prelude site. In fact, most of the alert ‘trimming’ that we did was in the LML rules which monitor log files.
An alert is the outcome of an event that is triggered by the occurrence of an anomaly that is being looked for. When we refer to an ‘alert’, technically it means a standardized representation of an event in IDMEF. So the role of the sensor is to convert the event into an XML-based IDMEF alert. The Prelude Manager receives these notifications and takes the alert structure and loads it into the database as tables. Through the web console, Prelude allows the user to view the alerts in a easily readable format as HTML table rows or in raw XML. Here is the entire ‘alert’ namespace for IDMEF used by Prelude.
Let’s look at two of our implementations for generating alerts. One is the Tattletale daemon which detects instances of root privilege escalation. The other is Continuous Mouse Biometric (CMB) alerts which detect when an unauthorized user is attempting to use a particular machine. Both detectors enter data into a log file when an event is encountered. Tattletale uses ‘syslog’ and CMB uses ‘messages’. Prelude LML monitors these log file by using PCRE configured in ‘pcre.rules’ which further breaks down the regex tests in specific rules files. Inside ‘pcre.rules’ we find the following lines that tell the PCRE engine when it finds a regex matching ‘ttd’ or ‘CMB’, look further into more specific rules files.
regex=ttd; include = ttd.rules;
regex=CMB; include = cmb.rules;
Here is ‘ttd.rules’ with an example line that will match in the log file. The CMB detection alert is very similar.
# Sep 19 09:58:55 pkirack3 ttd Output: Possible Root Escalation - host = pkirack3 pid=21319 proc=bash uid=24064 user=flans ancestors=(1 3709 21306 21309 213 11)
regex=Possible Root Escalation - host = ([^ ]+) pid=(\d+) proc=([^ ]+) uid=(\d+) user=([^ ]+) ancestors=\(((\d+\s?)+)\); \
classification.text=Possible Root Escalation; \
analyzer(0).class=Root Access; \
assessment.impact.description=Root access for User $5 with ancestors $6; \
A regex match will generate an IDMEF alert with text ‘Possible Root Escalation’ along with other useful alert parameters. Here is how the TTD alert ultimately appears in Prewikka.
Installing and Building Sensors
Installing and configuring the sensors turned out to be a bit tricky. Through trial and error we eventually got it right. Obviously, the component versions needed to be compatible and it was not always clear. We found a distribution for version .9 and stayed with it. The detailed steps for installation and configuration and package information for Snort and Prelude LML are in Sensor Installation. Basically, there are a lot of dependencies which creates a lot of compatibility problems. But once that is resolved, configuration and initialization are fairly straightforward. We had to build Snort with Prelude support because the version we need was not a Debian package. We will discuss this after an overview of the basic process.
- install the sensor noting dependencies
- run adduser for the sensor and manager to set up the sensor-manager SSL connection
- edit the appropriate configuration file to set the manager address
- start up the sensor with parameters – create scripts appropriately for restart upon rebooting
for example, with script… /etc/init.d/prelude-lml start
and without script… snort -D -c /usr/share/prelude/sensor-dist/snort-2.4.5/etc/rules/snort.conf -i eth0
- check prewikka on the prelude manager to make sure that the sensor shows up and is online.
Now we should take a quick look at sensor creation from scratch. As said before, a Prelude sensor can be built using C, Perl or Python. The key is building a Prelude client object which is the interface with Libprelude. We created a simple sensor using C using the following code. This is a great example of how straightforward developing for the Prelude API is. Building a Prelude Sensor using C, Perl or Python. Here are the basic steps for all three methods.
- initialize the Prelude library
- create the prelude_client_t
- start the client
- initialize manager connection
- send the initial heartbeat
- initialize the heartbeat timer flags
- use the low or high level API
- the low level API is faster and more optimized - but more error prone
- sending the actual IDMEF message
- destroy the client
Attack correlation became a central issue in our research. We found that in order to react to an attack effectively, one needs to have more than just alerts. Usually effective attacks are preceded by an orchestrated pattern of some sort that can be difficult to identify. The strength of using a hybrid IDS really became apparent at this point because we had alert data from network and host-based sensors giving us a clearer view of what might be happening. Often one or the other is just not enough information. But we still had the difficulty of deciding what we needed to look for, whether we were getting the crucial alerts and in general sorting through many attack patterns and schemes. This was even further complicated by actual implementation. At time we started working on Mithril, Prelude had no ability to perform correlations. With our first pass we decided to do some simple correlations based on alert types and time intervals. This allowed us to detect basic scanning attacks and brute force attempts on an SSH account. The brute force attack could detected by setting a threshold rate of alert generation. This was performed by using Prelude alert filters and applying the appropriate time interval for the database query. The source IP was set as a query parameter as well. This was done our implementation of Prelude version .8. Here is a diagram that shows how we modified the Piwi interface to indicate the attack. Our system policy response screen is also displayed. At this point we did not have an automated response for specific attacks in place.
With version .9, Prelude provided a way of performing correlations using the standard SEC (Simple Event Correlator). It was developed as an initial attempt to provide correlation for Prelude. For future versions, the Prelude team plans to provide built-in correlation tools integrated directly into the Prelude Manager. But in the meantime, using SEC provided us with a very effective and straightforward means to perform correlations. Its design is very similar to that of Prelude LML. The PCRE engine loads rulesets that perform regex comparisons on input streams. In order to make Prelude work with SEC, all of the generated alerts were intercepted from the Prelude Manager as IDMEF and the Prelude team built a conversion module for SEC that allowed it to scan IDMEF as its input. Prelude provided some simple rulesets, but we developed a few more to suit or needs. Here are some of the basics. The SEC engine is a Perl script which uses a file stream as input and loads rulesets at runtime. A ruleset is an ordered set of rules stored in a .sec text file. A rule is built using the following key-value pairs. More information is available on SEC Basics and Resources.
rule type (9 supported )
pattern type (13 supported / customizable)
rule triggering pattern
identifying name of the correlation
exit at this rule or continue to next
action list upon success
time window in seconds
for Calendar type
Here is a useful diagram that shows how Prelude and SEC are integrated. It also indicates how similar the roles of SEC and Prelude LML are in terms of scanning input data for pattern matching. On the left, we see the sensor inputs for the NIDS and the HIDS. All of the alerts are fed into the Prelude Manager where they are loaded into the database and also sent to SEC. In actuality, SEC is acting as a sensor as well. Where the input data for Snort and Prelude LML are packets and logfiles respectively, all of the alerts as an IDMEF stream is the input for SEC.
We did some research on correlation techniques and theories and we came up with some general guidelines about what we need to implement in order to get useful information about attacks. This diagram of a typical attack graph was an early attempt to determine what we needed in terms of correlations. Terminology was often a source of confusion, so we agreed upon some common concepts. Alerts are generated by the sensors. Correlation filters are essentially templates that wait for alerts to trigger instantiations as active correlations. Once we have an active correlation, further alerts that fit the correlation are prerequisites until it matures either as an attack or eventually times out. When a threshold is met - we have an actual attack. The names at the top show the corresponding terminology of SEC.
With Prewikka and Prelude version .9, we had our correlations mechanism in place. Once an SEC ruleset detects an actual attack, it creates a correlation alert and feeds it back to the Prelude Manager which in turn loads it into the database. The Prewikka alerts diagram shown earlier has a few examples of these Correlation Alerts generated by SEC. This diagram shows the system response table as well as the specific attack type responses. By changing the security level and hitting ‘enforce’, a series of scripts are fired which alter system-wide policy settings such as OTP for Single Sign-On and OTP for All Servers. The lower table shows the response associations with specific types of correlated attacks. We implemented six of them. The are brute force attempts on an SSH logon, excessive single user activity, a portscan followed by an SSH attempt from the same source IP address , root privilege escalation from a remote IP address, a continuous mouse biometric failure and an event scan where there is a single source IP scanning a host on the network. For each of these correlations, we developed a ruleset file for SEC to load at startup. Details about how this is done are on Building an SEC Ruleset for Attack Correlations. The script execution parameter can be set to manual or automatic and is set when the user hits ‘enforce’. These values are stored on the server. When SEC detects an attack, its ‘action’ field fires a script which checks the value before proceeding with a predetermined response. If set to manual, it does nothing. If automatic, it carries out the response. If manual, we can fire the response script manually from a drop-down menu which is tied to the Correlation Alert in the Prewikka alerts view. In this drop-down menu we also added inspection utilities such as ‘netstat’ and ‘nmap’ to complement the existing menu items ‘whois’ and ‘traceroute’. This was done to enhance the first level of inspection abilities tied to specific alerts. Details concerning the installation and configuration of Prewikka as well as our file modifications can be found at Prewikka Installation and Modifications.
Response Mechanisms and Adaptive System Policy
Our response framework is all script-based. For specific types of attacks, we used Perl scripts to perform such tasks as altering IP tables. The system-wide policy settings are enacted by the cfEngine configuration engine. At this point, we divided the adaptive system in five distinct levels of security as a proof of concept. Each level that is selected and enforced fires a script on the server that has preset parameters. This should be extended so that each level can be fine-tuned for specific system needs. The table shows general assumptions about what each level means, but its functionality can be greatly improved. In fact, the interface configurations should be stored as an XML policy file so that state can be maintained and modified. Each level should be linked to a configuration web page so that an administrator can alter system security settings efficiently.
One result we wanted to accomplish was to provide a useful framework for our security group. We met with them frequently in order to understand issues they dealt with on a daily basis. They use the Bro IDS which detects events and generates email messages as notifications. Then there is an inspection performed on the appropriate log files to determine the nature of the event that has occurred. This can be an extremely tedious process. The advantage of having a web interface like Prewikka is that troublesome alerts can be seen as they occur in one place where they can be visually correlated with other events that may be related. The interface acts as a central management console. From an IDS and forensic point of view, it serves as an entry point for the inspection process. We added typical initial actions for an alert hooked off of the IP address so that one could immediately run netstat and nmap as well as whois and traceroute. This functionality cuts down a lot of unnecessary, time-consuming activity navigating through a shell when all that is needed is routine inspection. The drawback of a web interface is the lack of flexibility as a security tool. Functionality is limited by script calls that are predetermined by the interface code. So the only feasible attack responses through the web interface are simple script calls that need specific parameters that are easily accessible. Everything else must be done by dropping into a shell. But as an attack inspection and a simple response tool, Prewikka can prove very useful to a security team. Also, overall system policy changes can be enacted very easily through the Mithril policy view.
Presently we are finishing up work on creating a sensor patch so that the Bro IDS acts as a Prelude sensor. It involves translating the Bro event notifications into IDMEF and sending messages to the Prelude Manager. The motivation behind this came from the fact that out security team uses Bro as its IDS and reacts to email notifications when events are triggered. This will bring their detection and response issues into the Prelude system for first inspection of possible attacks. Our goal has been to centralize management as much as possible to reduce inspection time without compromising present capabilities. The use of Prelude and our correlation and policy tools should be an enhancement to the current state and not a replacement. We also need to continue work on increasing the effectiveness and reliability of cfEngine for response scenarios. This will greatly determine the necessary parameters for making the adaptive policies more fine-grained and tunable. Forensic and analysis tools employed after a compromise need to be addressed as well. Then there are a number of issues that depend upon scaling up the actual testbed. We will need to implement the Mithril system on a sizable, dedicated network so that we can benchmark the responsiveness and isolate its inherent weaknesses. This will require the ability to perform orchestrated attacks for real-world simulations. These tests will give us a reasonable idea of how well we can detect many different types and scales of attacks. And once the system is compromised, we need to determine how well Mithril reacts by containing the damage, adapting to the new conditions and finally returning to the normal state once everything has been resolved.
This work was supported by the Office of Naval Research (ONR) grant N0001404-1-0562. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the view of ONR.