-----BEGIN PGP SIGNED MESSAGE-----
MIT krb5 Security Advisory 2005-002
Original release: 2005-07-12
Topic: buffer overflow, heap corruption in KDC
The MIT krb5 Key Distribution Center (KDC) implementation can corrupt
the heap by attempting to free memory at a random address when it
receives a certain unlikely (but valid) request via a TCP connection.
This attempt to free unallocated memory can result in a KDC crash and
consequent denial of service. [CAN-2005-1174, VU#259798]
Additionally, the same request, when received by the KDC via either
TCP or UDP, can trigger a bug in the krb5 library which results in a
single-byte overflow of a heap buffer. Application servers are
vulnerable to a highly improbable attack, provided that the attacker
controls a realm sharing a cross-realm key with the target
realm. [CAN-2005-1175, VU#885830]
An unauthenticated attacker may be able to use these vulnerabilities
to execute arbitrary code on the KDC host, potentially compromising an
entire Kerberos realm. No exploit code is known to exist at this
time. Exploitation of these vulnerabilities is believed to be
An unauthenticated attacker may be able to execute arbitrary code on
the KDC host, potentially compromising an entire Kerberos realm. An
unsuccessful attack against the heap corruption vulnerability may
result in a denial of service by crashing the KDC process.
* [CAN-2005-1174] affects the KDC implementation in all MIT krb5
releases supporting TCP client connections to the KDC. This
includes krb5-1.3 and later releases, up to and including
* [CAN-2005-1175] affects KDC implementations and application servers
in all MIT krb5 releases, up to and including krb5-1.4.1.
Third-party application servers which use MIT krb5 are also
* The upcoming krb5-1.4.2 release will have fixes for these
* WORKAROUNDS: Disabling TCP support in the KDC avoids one
vulnerability [CAN-2005-1174]. The single-byte overflow
[CAN-2005-1175] is still possible even without KDC TCP support
enabled. Running the KDC from init or from some similar automatic
respawning facility may reduce the durations of denials of service,
but this approach may make it difficult to detect deliberate attacks
targeted at code execution.
* Apply the patch at:
The associated detached PGP signature is at:
The patch was generated against the krb5-1.4.1 release. It may
apply, with some offset, to earlier releases. On releases prior to
krb5-1.3, only the patch to lib/krb5/krb/unparse.c should be
This announcement and related security advisories may be found on the
MIT Kerberos security advisory page at:
The main MIT Kerberos web page is at:
Thanks to Daniel Wachdorf for reporting these vulnerabilities.
Kerberos 5 principal names may have an arbitrary number of components.
The krb5_unparse_name() function in the MIT krb5 library converts an
internal representation of a Kerberos principal name into a
human-readable string. The internal representation might have
originated from the decoding of a Kerberos protocol message.
The single-byte overflow occurs whenever the krb5_unparse_name()
function is called on a principal name having zero components. The
function writes a null byte to an address one beyond the end of a
buffer allocated my malloc(). The corresponding krb5_parse_name()
function never generates an internal representation having zero
components; instead, it generates at least one zero-length component.
The current string representation form of Kerberos principal names has
some ambiguity between a zero-component principal name and a
one-component principal name having a zero-length single component.
Application servers which call krb5_unparse_name(), directly or
indirectly, are vulnerable to the single-byte overflow in
krb5_unparse_name(), provided that the attacker controls a realm which
shares a cross-realm key with the target realm. This enables the
attacker to use a cross-realm ticket for a zero-component client
principal name, which the application server will then pass to
krb5_unparse_name(), triggering the single-byte overflow.
For this attack to succeed, the attacker needs access to a KDC in the
target realm which will create a ticket for a zero-component client
principal name. Since the current MIT krb5 KDC implementation will
refuse to create such a ticket, the attack is unlikely to succeed
unless the implementation has been altered to allow the issuance of
tickets for zero-component client principal names.
When the KDC fails to find the principal with a zero-component name in
its database (such a principal is very unlikely to exist in most
databases, as there are extremely few uses for such a principal), it
attempts to encode an error packet containing the offending principal
name, using prepare_error_as() or prepare_error_tgs(). This encoding
attempt fails inside encode_krb5_error(), since the ASN.1 encoder
function asn1_encode_principal_name() interprets the internal
representation of a zero-component principal name as an error
encode_krb5_error() does not allocate an output buffer when it
encounters an error condition. While the UDP request handling code in
kdc/network.c:process_packet() does not attempt to free the output
buffer containing the encoded message when it encounters an error, the
TCP request handling code in process does free the buffer inside
kill_tcp_connection(), which attempts to free unallocated memory
pointed to by an uninitialized pointer.
2005-05-12 original release
Copyright (C) 2005 Massachusetts Institute of Technology
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.5 (SunOS)
-----END PGP SIGNATURE-----
kerberos-announce mailing list
|Free forum by Nabble||Edit this page|