text
stringlengths
4
429
Query Type
Others
15 (MX)
143,015
1 (A)
52,787
28 (AAAA)
12,254
Table 4: Breakdown of qtype by country code of recursive, for all five studied Aurora botnet CnCs.
Highlights: (a) Only the US victims were compelled to perform MX queries (qtype 15); all networks in
China and other countries never performed an MX query; (b) No AAAA (qtype 28) queries were
performed by international victims, who were presumably collateral victims; the pairing of AAAA to A
queries is discussed below; and (c) Most queries were MX (68% overall), and the attack heavily biased
towards the US (also 68 % overall).
CnC Domains over Time
Damballa
s analysis of DNS data has revealed the very early origin (July 2009) of the botnet. Even
during this early deployment, the botnet was widely dispersed. Since these were the first DNS
resolutions for these attacks, it is reasonable to assume they are associated with the botmaster (e.g.,
testing or configuring their attack), and not victims. Thus, these resolutions might correspond to
several CnC sites. If this theory is correct, it suggests that, despite using
naive
DDNS services typical of
novice botnet operators, the Aurora botmasters had considerable resources available to them.
CnC_Domain.1
The first resolution for CnC_Domain.1 came from within Google China
s offices. It was followed
hours later by resolutions inside Google
s offices in Mountain View, California. The pattern of lookups is
remarkable, and is worth closer study. The first queries for CnC_Domain.1 were:
2009-12-16 05:26:44
2009-12-17 22:39:09
2009-12-17 22:39:09
AAAA
AAAA
(Google China)
(Google Mountain View)
(Google Mountain View)
Page 14
The Command Structure of the Aurora Botnet
Counting Attempted Exfiltration Events
Other patterns of DNS messages in CnC_Domain.1 suggest the attempted exfiltration of data.
Consider this ordering of queries:
2009-12-18 06:29:09
2009-12-18 06:29:09
(Google Mountain View)
(Google Mountain View)
The queries both happen in under a second, indicating that a host using a recursive resolver wished to
send email to the CnC_Domain.1 CnC (hence the MX lookup). Dynamic DNS zones, however, almost
never have valid MX RRsets, or if they do they are pointed to abuse@traps or spamtraps. Only a few
DDNS providers offer mail, and the first query was therefore answered with an empty record
(NOERROR status, with zero answers). As a result, the victim immediately performed an A query, to use
the IP address for email. Whether these queries were followed by actual or successful email events is
not known.
All MX queries in CnC_Domain.1 came from the United States (and no other network outside the
US performed such a query before the news broke). Before January 10th 2010, some 110,810 MX
queries came from Google Mountain View, and one came from Comcast (San Jose). This Comcastbased query may have been testing by a Google security engineer, or it may have been an infection on
a notebook after work (since the query took place in the late evening hours, PST).
From the volume of messages, it is presumed that each MX query corresponds to a single email
exfiltration attempt. It would be hard to imagine a botmaster being able to direct these events
individually. Thus, it may not be the case that bots were instructed to email materials when a specific
event took place. Or the consistent pattern of queries could be the result of persistent searches of a
hard drive, and attempted, periodic exfiltration of any useful data. This conclusion is speculation.
The lack of any diurnal pattern to these events does indicate that the trigger event for an MX lookup
was not human-driven (e.g., the arrival of email on a victim machine, or selected actions by the
botmaster). It is not known what information was taken, if any, or if these queries were in fact victim
behavior. Public accounts from Google indicate that the attackers sought email records of human
rights activists.
It is speculated that Google would have prevented the
direct-to-MX
behavior of hosts within their
network. That is, in many corporate networks, individual user machines are prohibited from sending
email directly, and must instead use a smart host or authenticated relay system. Thus, these MX
lookups may well be side effect of an unsuccessful exfiltration effort. The malware also used ports 443
and 8585 for CnC, and could be instructed to perform any command.
CnC_Domain.3
The CnC_Domain.3 CnC domain is interesting because of its age. The botnet dates back to July 14,
2009, fell dormant for months, and then became active again within Google
s network. Of the five CnC
domains studied in detail within this report, this is the oldest, and most strongly suggests an origin for
the botnet.
The early queries for the Aurora CnC domain CnC_Domain.3 took place in the HangZhou region,
with some occurring in Beijing. The domain had a remarkable number of queries from mainland China
Page 15
The Command Structure of the Aurora Botnet
and collocation facilities in the US within minutes of being created. Seconds later, another query came
from Chinanet
s network in the Chongqing area. The close timing of these suggests the owners of
CnC_Domain.3 had access to ISP, university, and commercial transit.
2009-07-14 02:50:03
2009-07-14 02:57:38
2009-07-14 02:58:31
2009-07-14 03:03:11
2009-07-14 03:03:44
2009-07-14 03:04:28
2009-07-14 03:13:18
(HiNet Taiwan)
(CHINANET Jiangsu)
(CHINANET HangZhou)
(HangZhou Institute of Electronic Engineering)