1. JOHN P. JOHN FANG YU YINGLIAN XIE MARTÍN ABADI ARVIND KRISHNAMURTHY PRESENTATION BY SAM KLOCK Searching the Searchers with SearchAudit

2. Motivation We can find this via a Google search

3. Motivation (cont’d) Search engines open opportunities for attackers  Construct clever queries  Find vulnerable sites  Plant malware; spam (e.g., MyDoom)  Do so stealthily and cheaply Mitigation strategy: identify malicious queries  May be able to deny results to user  Identify attackers (probably bots)  Interpret strategy, then anticipate and prevent The question: how to do so

4. Proposed Approach SearchAudit  Framework for generating malicious queries Input:  Seed set of known malicious queries  Search logs Output:  Large set of suspicious queries  Regular expressions matching queries inurl:gotoURL.asp?url= filetype:asp inurl:"shopdisplayprod ucts.asp" ext:pl inurl:cgi intitle:"FormMail *" -"*Referrer" -"* Denied" -sourceforge -error -cvs -input filetype:cgi inurl:tseekdir.cgi... SearchAudit inurl:gotoURL.asp?url= filetype:asp inurl:"shopdisplayprod ucts.asp" ext:pl inurl:cgi intitle:"FormMail *" -"*Referrer" -"* Denied" -sourceforge -error -cvs -input filetype:cgi inurl:tseekdir.cgi... inurl:gotoURL.asp?url= filetype:asp inurl:"shopdisplayprod ucts.asp" ext:pl inurl:cgi intitle:"FormMail *" -"*Referrer" -"* Denied" -sourceforge -error -cvs -input filetype:cgi inurl:tseekdir.cgi... inurl:gotoURL.asp?url= filetype:asp inurl:"shopdisplayprod ucts.asp" ext:pl inurl:cgi intitle:"FormMail *" -"*Referrer" -"* Denied" -sourceforge -error -cvs -input filetype:cgi inurl:tseekdir.cgi... "/includes/joomla\.php " site:\.[a-zA- Z]{2,3} "/includes/class_item\.php" site:[^?=#+@;&:]{2, 4} "php-nuke" site:[^?=#+@;&:]{2, 4} "modules\.php\?op=modl oad" site:\.[a-zA- Z0-9]{2,6} Seed setSearch logs Expanded setRegular expressions

5. Proposed Approach (cont’d) Needed to implement:  Seed set: milw0rm.com  Search logs: Microsoft Research  Bing  Way to expand seed set into more queries  Way to infer regular expressions Intended benefits:  Harvesting lots of information  Three months: ~1.2 TB of logs  Interpret relationship between queries and attacks  Use queries to find potential victims  Stop attacks

6. SearchAudit Query identification Query analysis

7. Query Identification: Expansion Basic idea: bootstrap on seed set  Search logs for exact matches to seed queries  Record IPs of hosts making seed queries  Add other queries from those IPs to set  Intuition: make one malicious query, will probably make more Account for DHCP Seed queries IP addresses Queries made by IPs Log search Queries made on same day

8. Query Identification: Regular Expressions Goals:  Account for variation in queries  Take advantage of scripting See paper for generation algorithm Compute score for generated expressions  Lower score: more specific  Goal: discard overly general expressions (score > 0.6) Consolidate to avoid overlap Avoid proxies, public NAT for performance Loopback for more queries

9. Query Identification: Results Data from Bing and milw0rm  500 queries  Logs for Feb. 2009, Dec. 2009, Jan. 2010  ~2 billion views per month System implemented on Dryad/DryadLINQ Initial observations:  Using specificity scores < 0.6 seems to be effective  Based on cookie heuristic  Proxy elimination does not limit results

10. Query Identification: Results (cont’d) Query expansion:  122 of 500 queries matched in logs: 174 unique IPs  Expanded to 800 unique queries, 264 IPs  Regular expressions matched 3,560 queries, 1,001 IPs Incomplete seeds  Tried with subsets of original set  Coverage still good

11. Query Identification: Results (cont’d) Loopback:  Multiple loopbacks got more results  One iteration is good enough Overall statistics  10,000s IPs each month  100,000s unique queries each month  Dec. 09: set of unusual attacker IPs cause spike

12. Query Identification: Verification Want to show queries are malicious  Sometimes easy: 73% of queries associated with security/hacker sites  What about others? No ground truth exists So: look for bot-like features  Individual level (one IP)  Group level (multiple IPs) Individual bots  New cookie  Whether a link was clicked Groups of bots  Data often fixed by botnets  User agent string  Metadata for requests  Tendencies dictated by scripts  Pages viewed per query  Time between queries

13. Query Identification: Verification (cont’d) Substantial variation between host behavior for normal queries and suspicious queries

14. Observations on Stage One Regular expressions can become obsolete  Just need fresh logs and a new seed to get new ones Attacker awareness of technique yields adaptation  Example: mix in normal user queries  Goal: trick SearchAudit into identifying as proxy  Hard to do: needs to be appropriate to time and place  Anyway: proxy elimination is optimization only  Injecting randomness also possible, but makes querying less productive  Could obviate cookie heuristic, but it is replaceable All attackers need to be careful to succeed

15. Query Analysis

16. 42,000 IPs gave suspicious queries globally  U.S., Russia, China contribute almost 50%  10% of IPs gave 90% of queries Found 200 regular expressions Reveal three kinds of attack-related queries:  Vulnerable web sites  Forum spamming  Phishing on Windows Live Messenger

17. Queries for Vulnerable Websites Queries look for exploitable server vulnerabilities  GET variables embedded in URL (for SQL injection)  Server software with known vulnerabilities (e.g., status pages) SearchAudit as a defense:  Pull suspicious queries for vulnerabilities  Run queries; gather results  Inspect results for vulnerabilities  Notify sites of vulnerabilities inurl:index.php?content=X http://www.example.com/ind ex.php?content=X’%20OR%20’ 1’%20OR%20‘1=1’

18. Queries for Vulnerable Websites (cont’d) With identified queries:  Sampled 5,000 queries  Obtained 80,490 URLs from 39,475 sites Compared to malware/phishing lists:  3-4% on anti-phishing lists  1.5% on anti-malware lists SQL injection vulnerability:  Add a single-quote to variable in URL  Look for SQL error  12% of examined URLs showed an error

19. Queries for Forum Spamming Query motivation:  Find scriptable forums  Good for spam, PageRank Found 46 applicable regular expressions Most IPs show transient behavior: probably bots  All regular expression groups show at least one group similarity feature IPs got less aggressive over time: more stealthy

20. Queries for Forum Spamming (cont’d) Validation  Project Honey Pot  Dynamically generate e- mail address for each visiting IP  E-mail received: must be spam  12% of all IPs listed (vs. 0.5% for normal IPs) Applications  Use queries to find and clean targeted pages  Deny results to malicious queries

21. Phishing via Windows Live Messenger Queries triggered by normal users  Victim receives message from a contact  Follow link for party photos  Taken to fake WLM login  After giving credentials, redirected to Bing search for “party” Bing search to avoid costs of hosting

22. Phishing via WLM (cont’d) Detect via query referral field (source page)  Found two regular expressions for referrals  Both expressions: victim username embedded in URL Over 180 phishing domains for 12 IPs detected Compromised accounts show different login behaviors

23. Conclusion Presented framework for finding suspicious queries  Input: search logs, small set of seed queries  Output: regular expressions, millions of suspicious queries Analyzed suspicious queries  Identified possible attacks  Suggested means of prevention Generally: attempted to demonstrate relationship between suspicious queries and the possibility of attack