string.c: do not interpret '\\' before '\0' (CVE-2019-15846)
[exim.git] / doc / doc-txt / cve-2019-15846 / qualys.mbx
1From Wed Aug 14 01:29:25 CEST 2019
2Date: Tue, 13 Aug 2019 23:29:25 +0000
3From: Qualys Security Advisory <>
4To: Heiko Schlittermann <>
5Subject: Re: Help evaluating a Bug in Exim MTA
6Return-Path: <>
7Authentication-Results:; iprev=pass
8 ( smtp.remote-ip=; spf=pass
9; dkim=pass header.s=qualyscom
10 header.a=rsa-sha256; dkim=pass
11 header.s=selector2-qualys-onmicrosoft-com header.a=rsa-sha256; dmarc=none
13Authentication-Results:; spf=pass
14Status: O
15Content-Length: 3899
16Lines: 80
18Hi Heiko,
20On Mon, Aug 12, 2019 at 11:56:12PM +0200, Heiko Schlittermann wrote:
21> So I'd say, you do not need to rush, but I'd like to close it sooner or
22> later in either manner.
24OK, below is our preliminary analysis. First:
26- From an attacker's point of view, most calls to
27 string_interpret_escape() are uninteresting. For example, nextitem()
28 in src/filter.c checks for buffer overflows, and string_dequote()
29 seems to process trusted strings only (strings from configuration
30 files).
32- On the other hand, string_unprinting() is very interesting:
34 - It is used in tls_import_cert() (for peercert, for example); but
35 certificates are in PEM format (i.e., base64) and hence unlikely to
36 contain the problematic backslash-null-byte sequence.
38 - It is used for peerdn and sni in src/spool_in.c; but peerdn is used
39 only if client certificates are processed by Exim, and this is not
40 the default (and although some sites use client-certificate
41 authentication, this is not very common, and hence not very
42 interesting for an attacker).
44 - In any case, as long as Exim supports and accepts tls connections,
45 an attacker can send an sni, and hence reach the problematic
46 string_unprinting() and string_interpret_escape() functions.
48Next question: is it possible to send an sni that is written to the
49spool header file and that ends with the problematic backslash-null-byte
50sequence? The answer is yes, because of what we believe is another bug,
51in string_printing(): the sni is written to the spool header file via
52string_printing(tls_in.sni), which escapes characters with backslash,
53but does *not* escape the escaping character itself (backslash),
54although it definitely should.
56This bug is what makes it possible to reach and trigger the bug in
57string_unprinting() and string_interpret_escape(), with an sni that ends
58in an unescaped backslash (followed by the terminating null byte).
60Last question: is this exploitable? The answer is, almost certainly, yes
61(and, because spool_read_header() runs as root, this means remote root).
63The sni is read from the spool via string_unprinting(string_copy()), and
64both string_unprinting() and string_copy() use store_get(): as a result,
65the destination buffer is allocated right after the source buffer, and
66the characters that are read out-of-bounds after the end of the source
67buffer are the first characters of the destination buffer, which are
68fully under the attacker's control. This results in a heap overflow
69whose length and contents are both under the attacker's control (we
70verified this). This is almost certainly exploitable.
72Our advice is to start the security-release process as soon as possible.
73We know it is very painful, but we are really confident that this bug is
74exploitable; we will try to confirm this in the next few days. We also
75believe that an Exim server must support and accept tls connections to
76be remotely exploitable (via sni).
78During our analysis of this bug, we probably spotted three other bugs:
80- The unescaped backslash in string_printing() that we mentioned above.
82- A bug in spool_read_header(): before the for (;;) loop, p is set to
83 big_buffer + 2; and inside the loop, big_buffer may be re-allocated;
84 but p is never updated. This can lead to a use-after-free (we did not
85 assess the security impact of this bug, though).
87- A bug in spool_write_header(): the return value of tls_export_cert()
88 is not checked (for ourcert, but more importantly, for peercert). If
89 this function fails (maybe because big_buffer is not big enough), then
90 big_buffer may be uninitialized or unterminated, and garbage may be
91 written to the spool file (we did not assess the security impact of
92 this bug, either).
94We are at your disposal for questions, comments, and further
95discussions. Thank you very much for reaching out! With best regards,
98the Qualys Security Advisory team
99From Mon Aug 19 00:23:03 CEST 2019
100Date: Sun, 18 Aug 2019 22:23:03 +0000
101From: Qualys Security Advisory <>
102To: Heiko Schlittermann <>
103Subject: Re: Help evaluating a Bug in Exim MTA
104Return-Path: <>
105Authentication-Results:; iprev=pass
106 ( smtp.remote-ip=; spf=pass
107; dkim=pass header.s=qualyscom
108 header.a=rsa-sha256; dkim=pass
109 header.s=selector2-qualys-onmicrosoft-com header.a=rsa-sha256; dmarc=none
111Authentication-Results:; spf=pass
112Status: RO
113Content-Length: 2484
114Lines: 59
116Hi Heiko,
118On Tue, Aug 13, 2019 at 11:29:25PM +0000, Qualys Security Advisory wrote:
119> we are really confident that this bug is exploitable
121We can confirm that this bug is indeed exploitable: we wrote a
122rudimentary exploit that remotely obtains root privileges (because
123deliver_message() runs as root).
125Some general notes on this exploit:
127- To the best of our knowledge, the string_interpret_escape() bug
128 (backslash-null) is remotely exploitable if and only if Exim supports
129 and accepts TLS connections (because the only attack vector that we
130 know of is the string_unprinting() of SNI).
132- Both OpenSSL and GnuTLS installations are exploitable.
134- Our exploit is Linux-specific (because our heap-overflow exploitation
135 is specific to glibc's malloc implementation), but works on both i386
136 and amd64.
138Some detailed notes on this exploit:
140- First, we connect to Exim with TLS and send an SNI that ends with
141 backslash-null (this SNI is written unmodified to the spool because of
142 the unescaped-backslash bug in string_printing2()).
144- Second, we exploit the backslash-null bug in string_interpret_escape()
145 (our SNI is read from the spool and unescaped by string_unprinting()),
146 and we transform this out-of-bounds read into an out-of-bounds write
147 (a heap overflow).
149- Next, we use this heap overflow to overwrite the header of a free
150 malloc chunk, and increase its size to make it overlap with other,
151 already-allocated malloc chunks.
153- Last, we allocate this enlarged malloc chunk, and use it to overwrite
154 large parts of the heap (the already-allocated malloc chunks) with
155 arbitrary data:
157 . we overwrite the "id" string: it is used to build the message-log
158 file name, and therefore allows us to write to "/etc/passwd" (by
159 overwriting "id" with "/../../../../../../../../etc/passwd");
161 . we overwrite the "sender_address" string: it is written to the
162 message-log file, and therefore allows us to add a new user to
163 "/etc/passwd".
165Other exploitation methods may exist. We will not publish our exploit:
166it is a quick and dirty proof of concept, and we will not have the time
167to clean it anytime soon. However, please feel free to quote us on the
168exploitability of this bug (we do have a working exploit), and please
169feel free to quote all or parts of this email in your announcements.
171We are at your disposal for questions, comments, and further
172discussions. Thank you very much! With best regards,
175the Qualys Security Advisory team