xenocara/doc/xorg-docs/specs/Xserver/analysis.xml
2015-05-10 10:24:37 +00:00

2235 lines
58 KiB
XML

<?xml version="1.0" encoding="UTF-8" ?>
<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook XML V4.3//EN"
"http://www.oasis-open.org/docbook/xml/4.3/docbookx.dtd"
[
<!ENTITY % defs SYSTEM "defs.ent"> %defs;
]>
<!-- by TeX4ht (http://www.cse.ohio-state.edu/~gurari/TeX4ht/)
xhtml,docbook,html,refcaption
-->
<book id="analysis">
<bookinfo>
<title>Analysis of the X Protocol for Security Concerns</title>
<pubdate>May 10, 1996</pubdate>
<authorgroup>
<author>
<firstname>David</firstname><othername>P.</othername><surname>Wiggins</surname>
<affiliation><orgname>X Consortium</orgname></affiliation>
</author>
</authorgroup>
<releaseinfo>X Version 11, Release &fullrelvers;</releaseinfo>
<releaseinfo>Draft Version 2</releaseinfo>
<copyright><year>1996</year><holder>X Consortium</holder></copyright>
<legalnotice>
<para>
Permission is hereby granted, free of charge, to any person obtaining
a copy of this software and associated documentation files (the
"Software"), to deal in the Software without restriction, including
without limitation the rights to use, copy, modify, merge, publish,
distribute, sublicense, and/or sell copies of the Software, and to
permit persons to whom the Software is furnished to do so, subject to
the following conditions:
</para>
<para>
The above copyright notice and this permission notice shall be included
in all copies or substantial portions of the Software.
</para>
<para>
THE SOFTWARE IS PROVIDED &ldquo;AS IS&rdquo;, WITHOUT WARRANTY OF ANY KIND, EXPRESS
OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE X CONSORTIUM BE LIABLE FOR ANY CLAIM, DAMAGES OR
OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
OTHER DEALINGS IN THE SOFTWARE.
</para>
<para>
Except as contained in this notice, the name of the X Consortium shall
not be used in advertising or otherwise to promote the sale, use or
other dealings in this Software without prior written authorization
from the X Consortium.
</para>
<para>X Window System is a trademark of The Open Group.</para>
</legalnotice>
<abstract>
<para>
This paper attempts to list all instances of certain types of security
problems in the X Protocol. Issues with authorization are not addressed.
We assume that a malicious client has already succeeded in connecting,
and try to assess what harm it can then do. We propose modifications to
the semantics of the X Protocol to reduce these risks.
</para>
</abstract>
</bookinfo>
<chapter id='Definition_of_Threats'>
<title>Definition of Threats</title>
<para>
We analyze the X protocol for the following threats.
</para>
<variablelist>
<varlistentry>
<term>Theft</term>
<listitem>
<para>
occurs when a client gains access to information owned by another client
without explicit permission from that other client. For this analysis,
we take a broad view of ownership: any information that exists in the
server due to the actions of a client is considered owned by that client.
Furthermore, the client that has input focus owns keyboard events, and the
client that owns the window that the pointer is in owns mouse events. This
view may reveal certain instances of "theft" that we don't care to stop,
but we think it is better to identify all potential candidates up front and
cull the list later than to do a partial analysis now and plan on reanalyzing
for remaining holes later.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>Denial of service</term>
<listitem>
<para>
occurs when a client causes another client or the user to lose the ability
to perform some operation.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>Spoofing</term>
<listitem>
<para>
occurs when a client attempts to mimic another client with the hope that
the user will interact with it as if it really were the mimicked client. A
wide variety of requests may be used in a spoofing attack; we will only
point out a few likely candidates.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>Destruction</term>
<listitem>
<para>
occurs when a client causes another client to lose information in a way
that the client or user is likely to notice. (This does not count expected
forms of destruction, e.g., exposures.)
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>Alteration</term>
<listitem>
<para>
occurs when a client causes another client to lose information in a way
that the client or user is unlikely to notice. e.g., changing one pixel in a
drawable.
</para>
</listitem>
</varlistentry>
</variablelist>
<para>
The line between alteration and destruction is subjective. Security
literature does often distinguish between them, though not always
explicitly. Alteration is often considered more insidious because
its effects may not be realized until long after it has occurred. In
the intervening time, each time the altered data is used, it can cause
more damage.
</para>
</chapter>
<chapter id='General_security_concerns_and_remedies'>
<title>General security concerns and remedies</title>
<para>
The following sections discuss security problems intrinsic to the X
Protocol. A statement of each problem is usually followed by potential
remedies. A few words here about possible remedies will help frame the
specific ones described below.
</para>
<para>
If a client attempts a threatening operation, the server may take one of
the following actions, listed roughly in order of severity:
</para>
<orderedlist>
<listitem>
<para>
Execute the request normally. This is the right choice when we decide that a
particular threat is not serious enough to worry about.
</para>
</listitem>
<listitem>
<para>
Execute the request in some modified form, e.g., substitute different
values for some of the request fields, or edit the reply.
</para>
</listitem>
<listitem>
<para>
Arrange to ask the user what to do, given some subset of the other choices
in this list. This must be used sparingly because of the performance impact.
</para>
</listitem>
<listitem>
<para>
Treat the request as a no-op. If the client will probably not notice, or
if it seems likely that the intent was benign, this is a good choice.
</para>
</listitem>
<listitem>
<para>
Send a protocol error to the client. If the client will be confused enough
by the other options that it will probably crash or become useless anyway,
or if it seems likely that the intent was malicious, this is a good choice.
</para>
<para>
Kill the client. This might be the right action if there is no doubt
that the client is hostile.
</para>
</listitem>
</orderedlist>
<para>
In most cases we present the one option that seems most appropriate to
counter the threat, taking into account the seriousness of the threat, the
implementation difficulty, and the impact on applications. Our initial bias
is to err on the side of stronger security, with the accompanying tighter
restrictions. As we uncover important operations and applications that the new
restrictions interfere with, we can apply selective loosening to allow
the desired functionality.
</para>
<para>
In some cases we will suggest returning an Access error where the X protocol
does not explicitly allow one. These new Access errors arise when a client
can only perform a (non-empty) subset of the defined operations on a
resource. The disallowed operations cause Access errors. The resource at
issue is usually a root window.
</para>
<sect1 id='Access_to_Server_Resources'>
<title>Access to Server Resources</title>
<para>
The X protocol allows clients to manipulate resources (objects) belonging to
other clients or to the server. Any request that specifies a resource ID is
vulnerable to some of the above threats. Such requests also provide a way for a
client to guess resource IDs of other clients. A client can make educated
guesses for possible resource IDs, and if the request succeeds, it knows it has
determined a valid resource ID. We call this "resource ID guessing" in the list
below.
</para>
<para>
One likely defense against these problems is to have the server send an
appropriate protocol error to deny the existence of any resource specified
by a client that doesn't belong to that client. A variation on this policy
lets cooperating groups of clients access each other's resources, but not
those of other groups. The Broadway project will initially use a less general
form of this idea by having two groups, trusted and untrusted. Trusted
clients can do everything that X clients can do today. They will be protected
from untrusted clients in ways described below. Untrusted clients will
not be protected from each other. Though this will be the initial design,
we need to make sure there is a growth path to multiple (more than two)
groups.
</para>
<para>
Most of the time, applications never access server resources that aren't
their own, so the impact of disallowing such accesses should be minimal.
There are a few notable exceptions, most of which will be discussed under
the relevant protocol requests. They are: ICCCM selection transfer, Motif
drag and drop, and server-global resources like the root window and default
colormap. Another major exception is the window manager, which routinely
manipulates windows of other applications. The solution for window managers
is to always run them as trusted applications.
</para>
<para>
The implementation difficulty of limiting access to resources should not
be large. All resource accesses eventually funnel down to one of two
functions in &lt;dix/resource.c&gt;:
<function>LookupIDByType</function> and
<function>LookupIDByClass</function>. A few lines of checking at
the top of these functions will form the heart of this defense. There is a
small problem because these functions are not told which client is doing the
lookup, but that can be solved either by adding a client parameter (probably
as a new function to preserve compatibility), or by using the server global
requestingClient.
</para>
<note>
<para>
ISSUE: are we really going to be able to get away with hiding trusted
resources, or will things like Motif drag and drop force us to expose
them? (Either way, the operations that untrusted clients can do to
trusted resources will have to be limited.) Is there something in Xt
or the ICCCM that breaks if you hide resources?
</para>
</note>
</sect1>
<sect1 id='Denial_of_Service'>
<title>Denial of Service</title>
<sect2 id='Memory_Exhaustion'>
<title>Memory Exhaustion</title>
<para>
Any request that causes the server to consume resources (particularly
memory) can be used in a denial of service attack. A client can use such
requests repeatedly until the server runs out of memory. When that
happens, the server will either crash or be forced to send Alloc errors.
The most obvious candidates are resource creation requests,
e.g., CreatePixmap, but in reality a large percentage of requests cause
memory allocation, if only temporarily, depending on the server
implementation. For this reason, the list of requests subject to this
form of denial of service will be necessarily incomplete.
</para>
<para>
To address this form of denial of service, the server could set
per-client quotas on memory consumption. When the limit is surpassed,
the server could return Alloc errors. The application impact is minimal
as long as the application stays within quota. The implementation difficulty
is another story.
</para>
<para>
Conceptually, it seems easy: simply have a way to set the limit, and on
every memory (de)allocation operation, update the client's current usage,
and return an error if the client is over the limit. The first problem
is something we've already touched on: the allocator functions aren't told
which client the allocation belongs to. Unlike resource lookups, allocations
are done in too many places to consider a new interface that passes the
client, so using the global requestingClient is practically mandatory.
</para>
<para>
The problems run deeper. The logical thing for the allocator to do if the
client is over its limit is to return NULL, indicating allocation failure.
Unfortunately, there are many places in the server that will react badly
if this happens. Most of these places, but not all, are "protected" by
setting the global variable Must_have_memory to True around the delicate
code. We could help the problem by skipping the limit check if
Must_have_memory is True. The best solution would be to bullet-proof the
server against allocation failures, but that is beyond the scope of Broadway.
Another consideration is that the additional checking may have a measurable
performance impact, since the server does frequent allocations.
</para>
<para>
A third problem is that there is no portable way to determine the size of
a chunk of allocated memory given just a pointer to the chunk, and that's
all you have inside <function>Xrealloc</function> and
<function>Xfree</function>. The server could compensate by recording the
sizes itself somewhere, but that would be wasteful of memory, since the
malloc implementation also must be recording block sizes. On top of that, the
redundant bookkeeping would hurt performance. One solution is to use a custom
malloc that has the needed support, but that too seems beyond the scope of
Broadway.
</para>
<para>
Considering all of this, we think it is advisable to defer solving the memory
exhaustion problem to a future release. Keep this in mind when you see quotas
mentioned as a defense in the list below.
</para>
</sect2>
<sect2 id='CPU_Monopolization'>
<title>CPU Monopolization</title>
<para>
Another general way that a client can cause denial of service is to flood
the server with requests. The server will spend a large percentage of its
time servicing those requests, possibly starving other clients and certainly
hurting performance. Every request can be used for flooding, so we will
not bother to list flooding on every request. A variation on this attack is
to flood the server with new connection attempts.
</para>
<para>
To reduce the effectiveness of flooding, the server could use a different
scheduling algorithm that throttles clients that are monopolizing the
server, or it could simply favor trusted clients over untrusted ones.
Applications cannot depend on a particular scheduling algorithm anyway,
so changing it should not affect them. The Synchronization extension
specifies a way to set client priorities, and a simple priority scheduler
already exists in the server to support it, so this should be simple to add.
</para>
</sect2>
</sect1>
</chapter>
<chapter id='Security_concerns_with_specific_window_attributes'>
<title>Security concerns with specific window attributes</title>
<sect1 id='Background_pixmap'>
<title>Background-pixmap</title>
<para>
Clients can use windows with the background-pixmap attribute set to None
(hereafter "background none windows") to obtain images of other windows. A
background none window never paints its own background, so whatever happened
to be on the screen when the window was mapped can be read from the
background none window with GetImage. This may well contain data from other
windows. The CreateWindow and ChangeWindowAttributes requests can set the
background-pixmap attribute set to None, and many window operations can cause
data from other windows to be left in a background none window, including
<function>ReparentWindow</function>,
<function>MapWindow</function>,
<function>MapSubwindows</function>,
<function>ConfigureWindow</function>, and
<function>CirculateWindow</function>.
</para>
<para>
Background none windows can also be used to cause apparent alteration.
A client can create a window with background none and draw to it. The drawing
will appear to the user to be in the windows below the background none window.
</para>
<para>
To remedy these problems, the server could substitute a well-defined
background when a client specifies None. Ideally the substituted background
would look different enough from other windows that the user wouldn't be
confused. A tile depicting some appropriate international symbol might be
reasonable. We believe that there are few applications that actually rely on
background none semantics, and those that do will be easy for the user to
identify because of the distinctive tile. Implementation should not be a
problem either. Luckily, the window background cannot be retrieved through
the X protocol, so we won't have to maintain any illusions about its value.
</para>
<note>
<para>
ISSUE: Some vendors have extensions to let you query the window background. Do
we need to accomodate that?
</para>
</note>
<note>
<para>
ISSUE: Will this lead to unacceptable application breakage? Could the server be
smarter, only painting with the well-defined background when the window actually
contains bits from trusted windows?
</para>
</note>
</sect1>
<sect1 id='ParentRelative_and_CopyFromParent'>
<title>ParentRelative and CopyFromParent</title>
<para>
Several window attributes can take on special values that cause them to
reference (ParentRelative) or copy (CopyFromParent) the same attribute from
the window's parent. This fits our definition of theft. The window
attributes are class, background-pixmap, border-pixmap, and colormap. All
of these can be set with <function>CreateWindow</function>; all but class
can be set with <function>ChangeWindowAttributes</function>.
</para>
<para>
These forms of theft aren't particularly serious, so sending an error
doesn't seem appropriate. Substitution of different attribute values seems
to be the only reasonable option, and even that is likely to cause trouble
for clients. Untrusted clients are already going to be prevented from
creating windows that are children of trusted clients (see CreateWindow
below). We recommend that nothing more be done to counter this threat.
</para>
</sect1>
<sect1 id='Override_redirect'>
<title>Override-redirect</title>
<para>
Windows with the override-redirect bit set to True are generally ignored by
the window manager. A client can map an override-redirect window that covers
most or all of the screen, causing denial of service since other applications
won't be visible.
</para>
<para>
To prevent this, the server could prevent more than a certain percentage
(configurable) the of screen area from being covered by override-redirect
windows of untrusted clients.
</para>
<para>
Override-redirect windows also make some spoofing attacks easier since the
client can more carefully control the presentation of the window to mimic
another client. Defenses against spoofing will be given under
<xref linkend='MapWindow' xrefstyle='select: title'/>
.
</para>
</sect1>
</chapter>
<chapter id='Security_concerns_with_specific_requests'>
<title>Security concerns with specific requests</title>
<para>
To reduce the space needed to discuss 120 requests, most of the following
sections use a stylized format. A threat is given, followed by an
imperative statement. The implied subject is an untrusted client, and the
object is usually a trusted client. Following that, another statement
starting with "Defense:" recommends a countermeasure for the preceding
threat(s).
</para>
<para>
Resources owned by the server, such as the root window and the default
colormap, are considered to be owned by a trusted client.
</para>
<sect1 id='CreateWindow'>
<title>CreateWindow</title>
<para>
Alteration: create a window as a child of another client's window, altering
its list of children.
</para>
<para>
Defense: send Window error. Specifying the root window as the parent will
have to be allowed, though.
</para>
<para>
Theft: create an InputOnly window or a window with background none on top
of other clients' windows, select for keyboard/mouse input on that window,
and steal the input. The input can be resent using SendEvent or an input
synthesis extension so that the snooped application continues to function,
though this won't work convincingly with the background none case because the
drawing will be clipped.
</para>
<para>
Defense: send an error if a top-level InputOnly window is created
(or reparented to the root). Countermeasures for background none and
SendEvent are discussed elsewhere.
</para>
<note>
<para>
ISSUE: The Motif drag and drop protocol creates and maps such a window (at
-100, -100, size 10x10) to "cache frequently needed data on window properties
to reduce roundtrip server requests." Proposed solution: we could only send an
error if the window is visible, which would require checking in, MapWindow,
ConfigureWindow, and ReparentWindow.
</para>
</note>
<para>
Theft: resource ID guessing (parent, background-pixmap, border-pixmap, colormap,
and cursor).
</para>
<para>
Defense: send Window, Pixmap, Colormap, or Cursor error.
</para>
<para>
Denial of service: create windows until the server runs out of memory.
</para>
<para>
Defense: quotas.
</para>
<para>
Also
<xref linkend='Security_concerns_with_specific_window_attributes' xrefstyle='select: title'/>
</para>
</sect1>
<sect1 id='ChangeWindowAttributes'>
<title>ChangeWindowAttributes</title>
<para>
Alteration: change the attributes of another client's window.
</para>
<para>
Theft: select for events on another client's window.
</para>
<para>
Defense for both of the above: send Window error.
</para>
<note>
<para>
ISSUE: The Motif drop protocol states that "the initiator should select for
DestroyNotify on the destination window such that it is aware of a potential
receiver crash." This will be a problem if the initiator is an untrusted
window and the destination is trusted. Can the server, perhaps with the help
of the security manager, recognize that a drop is in progress and allow the
DestroyNotify event selection in this limited case?
</para>
</note>
<note>
<para>
ISSUE: The Motif pre-register drag protocol probably requires the initiator
to select for Enter/LeaveNotify on all top-level windows. Same problem as
the previous issue.
</para>
</note>
<para>
Theft: resource ID guessing (background-pixmap, border-pixmap, colormap, and
cursor).
</para>
<para>
Defense: send Pixmap, Colormap, or Cursor error.
</para>
<para>
Also
<xref linkend='Security_concerns_with_specific_window_attributes' xrefstyle='select: title'/>
</para>
</sect1>
<sect1 id='GetWindowAttributes'>
<title>GetWindowAttributes</title>
<para>
Theft: get the attributes of another client's window.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
</sect1>
<sect1 id='DestroyWindow_DestroySubwindows'>
<title>DestroyWindow, DestroySubwindows</title>
<para>
Destruction: destroy another client's window.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
</sect1>
<sect1 id='ChangeSaveSet'>
<title>ChangeSaveSet</title>
<para>
Alteration: cause another client's windows to be reparented to the root when
this client disconnects (only if the other client's windows are subwindows of
this client's windows).
</para>
<para>
Defense: process the request normally. The trusted client gives away some of
its protection by creating a subwindow of an untrusted window.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='MapWindow'>
<title>MapWindow</title>
<para>
Spoofing: map a window that is designed to resemble a window of another client.
Additional requests will probably be needed to complete the illusion.
</para>
<para>
Defense:
</para>
<para>
We consider spoofing to be a significant danger only if the user is convinced
to interact with the spoof window. The defense centers on providing enough
information to enable the user to know where keyboard, mouse, and extension
device input is going. To accomplish this, the server will cooperate with the
security manager, an external process. The server will provide the following
facilities to the security manager:
</para>
<para>
1. A way to create a single window that is unobscurable by any window of any
other client, trusted or untrusted. It needs to be unobscurable so that it is
spoof-proof.
</para>
<note>
<para>
ISSUE: is a weaker form of unobscurability better? Should the window be
obscurable by trusted windows, for example?
</para>
</note>
<note>
<para>
ISSUE: does unobscurable mean that it is a child of the root that is always
on top in the stacking order?
</para>
</note>
<para>
2. A way to determine if a given window ID belongs to an untrusted client.
</para>
<para>
The security manager will need to select for the existing events
FocusIn, FocusOut, EnterNotify, LeaveNotify, DeviceFocusIn, and
DeviceFocusOut on all windows to track what window(s) the user's input is
going to. Using the above server facilities, it can reliably display the
trusted/untrusted status of all clients currently receiving input.
</para>
<note>
<para>
ISSUE: is it too much to ask the security manager to select for all these
events on every window? Do we need to provide new events that you select
for *on the device* that tell where the device is focused?
</para>
</note>
<para>
None of this should have any application impact.
</para>
<para>
The unobscurable window may be tricky to implement. There is already some
machinery in the server to make an unobscurable window for the screen saver,
which may help but may also get in the way now that we have to deal with two
unobscurable windows.
</para>
</sect1>
<sect1 id='Window_Operations'>
<title>Window Operations</title>
<para>
Specifically, ReparentWindow, MapWindow, MapSubwindows, UnmapWindow,
UnmapSubwindows, ConfigureWindow, and CirculateWindow.
</para>
<para>
Alteration: manipulate another client's window.
</para>
<para>
Theft: resource ID guessing (window, sibling).
</para>
<para>
Defense for both of the above: send a Window error unless it is a root
window, in which case we should send an Access error.
</para>
</sect1>
<sect1 id='GetGeometry'>
<title>GetGeometry</title>
<para>
Theft: get the geometry of another client's drawable.
</para>
<para>
Theft: resource ID guessing (drawable).
</para>
<para>
Defense for both of the above: send Drawable error. However, root windows
will be allowed.
</para>
</sect1>
<sect1 id='QueryTree'>
<title>QueryTree</title>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Window error.
</para>
<para>
Theft: discover window IDs that belong to other clients.
</para>
<para>
Defense: For the child windows, censor the reply by removing window IDs that
belong to trusted clients. Allow the root window to be returned. For the parent
window, if it belongs to a trusted client, return the closest ancestor window
that belongs to an untrusted client, or if such a window does not exist,
return the root window for the parent window.
</para>
<note>
<para>
ISSUE: will some applications be confused if we filter out the window manager
frame window(s), or other windows between the queried window and the root
window?
</para>
</note>
<note>
<para>
ISSUE: the Motif drag protocol (both preregister and dynamic) needs to be able
to locate other top-level windows for potential drop sites. See also
<xref linkend='Access_to_Server_Resources' xrefstyle='select: title'/>
.
</para>
</note>
</sect1>
<sect1 id='InternAtom'>
<title>InternAtom</title>
<para>
Theft: discover atom values of atoms interned by other clients. This lets you
determine if a specific set of atoms has been interned, which may lead to other
inferences.
</para>
<para>
Defense: This is a minor form of theft. Blocking it will interfere with many
types of inter-client communication. We propose to do nothing about this
threat.
</para>
<para>
Denial of service: intern atoms until the server runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='GetAtomName'>
<title>GetAtomName</title>
<para>
Theft: discover atom names of atoms interned by other clients. This lets you
determine if a specific set of atoms has been interned, which may lead to other
inferences.
</para>
<para>
Defense: This is a minor form of theft. We propose to do nothing about this
threat.
</para>
</sect1>
<sect1 id='ChangeProperty'>
<title>ChangeProperty</title>
<para>
Alteration: change a property on another client's window or one that was
stored by another client.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
<note>
<para>
ISSUE: Selection transfer requires the selection owner to change a property
on the requestor's window. Does the security manager get us out of this?
Does the server notice the property name and window passed in
ConvertSelection and temporarily allow that window property to be written?
</para>
</note>
<note>
<para>
ISSUE: should certain root window properties be writable?
</para>
</note>
<para>
Denial of service: store additional property data until the server
runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='DeleteProperty'>
<title>DeleteProperty</title>
<para>
Destruction: delete a property stored by another client.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
</sect1>
<sect1 id='GetProperty'>
<title>GetProperty</title>
<para>
Theft: get a property stored by another client.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
<note>
<para>
ISSUE: should certain root window properties be readable? Proposed answer: yes,
some configurable list. Do those properties need to be polyinstantiated?
</para>
</note>
<note>
<para>
ISSUE: Motif drag and drop needs to be able to read the following properties:
WM_STATE to identify top-level windows, _MOTIF_DRAG_WINDOW on
the root window, _MOTIF_DRAG_TARGETS on the window given in the
_MOTIF_DRAG_WINDOW property, and _MOTIF_DRAG_RECEIVER_INFO on
windows with drop sites. Additionally, some properties are needed that do not
have fixed names.
</para>
</note>
</sect1>
<sect1 id='RotateProperties'>
<title>RotateProperties</title>
<para>
Alteration: rotate properties stored by another client.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
</sect1>
<sect1 id='ListProperties'>
<title>ListProperties</title>
<para>
Theft: list properties stored by another client.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense for both of the above: send Window error.
</para>
<note>
<para>
ISSUE: should certain root window properties be listable?
</para>
</note>
</sect1>
<sect1 id='SetSelectionOwner'>
<title>SetSelectionOwner</title>
<para>
Theft: Steal ownership of a selection.
</para>
<para>
Denial of service: do this repeatedly so that no other client can own the
selection.
</para>
<para>
Defense for both of the above: have a configurable list of selections that
untrusted clients can own. For other selections, treat this request as a no-op.
</para>
<note>
<para>
ISSUE: how does the security manager get involved here? Is it the one that
has the configurable list of selections instead of the server?
</para>
</note>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='GetSelectionOwner'>
<title>GetSelectionOwner</title>
<para>
Theft: discover the ID of another client's window via the owner field of the
reply.
</para>
<para>
Defense: if the selection is on the configurable list mentioned above,
return the root window ID, else return None.
</para>
<note>
<para>
ISSUE: how does the security manager get involved here?
</para>
</note>
</sect1>
<sect1 id='ConvertSelection'>
<title>ConvertSelection</title>
<para>
Theft: this initiates a selection transfer (see the ICCCM) which sends
the selection contents from the selection owner, which may be another
client, to the requesting client.
</para>
<para>
Defense: since in many cases ConvertSelection is done in direct response
to user interaction, it is probably best not to force it to fail, either
silently or with an error. The server should rely on the security manager
to assist in handling the selection transfer.
</para>
<para>
Theft: resource ID guessing (requestor).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='SendEvent'>
<title>SendEvent</title>
<para>
A client can use SendEvent to cause events of any type to be sent to windows
of other clients. Similarly, a client could SendEvent to one of its own
windows with propagate set to True and arrange for the event to be propogated
up to a window it does not own. Clients can detect events generated by
SendEvent, but we cannot assume that they will.
</para>
<para>
Defense: ignore this request unless the event being sent is a ClientMessage
event, which should be sent normally so that selection transfer, Motif drag
and drop, and certain input methods have a chance at working.
</para>
<note>
<para>
ISSUE: does allowing all ClientMessages open up too big a hole?
</para>
</note>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='Keyboard_and_Pointer_Grabs'>
<title>Keyboard and Pointer Grabs</title>
<para>
Specifically, GrabKeyboard, GrabPointer, GrabKey, and GrabButton.
</para>
<para>
Denial of service/Theft: take over the keyboard and pointer. This could
be viewed as denial of service since it prevents other clients from getting
keyboard or mouse input, or it could be viewed as theft since the user
input may not have been intended for the grabbing client.
</para>
<para>
Defense: provide a way to break grabs via some keystroke combination, and
have a status area that shows which client is getting input. (See
<xref linkend='MapWindow' xrefstyle='select: title'/>
).
</para>
<para>
Theft: resource ID guessing (grab-window, confine-to, cursor).
</para>
<para>
Defense: send Window or Cursor error.
</para>
</sect1>
<sect1 id='ChangeActivePointerGrab'>
<title>ChangeActivePointerGrab</title>
<para>
Theft: resource ID guessing (cursor).
</para>
<para>
Defense: send Cursor error.
</para>
</sect1>
<sect1 id='GrabServer'>
<title>GrabServer</title>
<para>
Denial of service: a client can grab the server and not let go, locking
out all other clients.
</para>
<para>
Defense: provide a way to break grabs via some keystroke combination.
</para>
</sect1>
<sect1 id='QueryPointer'>
<title>QueryPointer</title>
<para>
Theft: A client can steal pointer motion and position, button input,
modifier key state, and possibly a window of another client with this request.
</para>
<para>
Defense: if the querying client doesn't have the pointer grabbed, and
the pointer is not in one of its windows, the information can be zeroed.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='GetMotionEvents'>
<title>GetMotionEvents</title>
<para>
Theft: steal pointer motion input that went to other clients.
</para>
<para>
Defense: ideally, the server would return only pointer input that was not
delivered to any trusted client. The implementation effort to do that
probably outweighs the marginal benefits. Instead, we will always return
an empty list of motion events to untrusted clients.
</para>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='TranslateCoordinates'>
<title>TranslateCoordinates</title>
<para>
Theft: discover information about other clients' windows: position,
screen, and possibly the ID of one of their subwindows.
</para>
<para>
Defense: send an error if src-window or dst-window do not belong to
the requesting client.
</para>
<para>
Theft: resource ID guessing (src-window, dst-window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='WarpPointer'>
<title>WarpPointer</title>
<para>
A client can cause pointer motion to occur in another client's window.
</para>
<para>
Denial of service: repeated pointer warping prevents the user from using
the mouse normally.
</para>
<para>
Defense for both of the above: if the querying client doesn't have the pointer
grabbed, and the pointer is not in one of its windows, treat the request as a
no-op.
</para>
<para>
Theft: resource ID guessing (src-window, dst-window).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='SetInputFocus'>
<title>SetInputFocus</title>
<para>
Theft: a client can use this request to make one of its own windows have
the input focus (keyboard focus). The user may be unaware that keystrokes
are now going to a different window.
</para>
<para>
Denial of service: repeatedly setting input focus prevents normal use of the
keyboard.
</para>
<para>
Defense for both of the above: only allow untrusted clients to
SetInputFocus if input focus is currently held by another untrusted client.
</para>
<note>
<para>
ISSUE: this will break clients using the Globally Active Input model
described in section 4.1.7 of the ICCCM.
</para>
</note>
<para>
Theft: resource ID guessing (focus).
</para>
<para>
Defense: send Window error.
</para>
</sect1>
<sect1 id='GetInputFocus'>
<title>GetInputFocus</title>
<para>
Theft: the reply may contain the ID of another client's window.
</para>
<para>
Defense: return a focus window of None if a trusted client currently has
the input focus.
</para>
</sect1>
<sect1 id='QueryKeymap'>
<title>QueryKeymap</title>
<para>
Theft: poll the keyboard with this to see which keys are being pressed.
</para>
<para>
Defense: zero the returned bit vector if a trusted client currently has
the input focus.
</para>
</sect1>
<sect1 id='Font_Requests'>
<title>Font Requests</title>
<para>
Specifically, OpenFont, QueryFont, ListFonts, ListFontsWithInfo, and
QueryTextExtents.
</para>
<para>
Theft: discover font name, glyph, and metric information about fonts that were
provided by another client (by setting the font path). Whether it is theft
to retrieve information about fonts from the server's initial font path
depends on whether or not you believe those fonts, by their existence in the
initial font path, are intended to be globally accessible by all clients.
</para>
<para>
Defense:
</para>
<para>
Maintain two separate font paths, one for trusted clients and one for untrusted
clients. They are both initialized to the default font path at server reset.
Subsequently, changes to one do not affect the other. Since untrusted clients
will not see font path elements added by trusted clients, they will not be
able to access any fonts provided by those font path elements.
</para>
<para>
Theft: resource ID guessing (font) (QueryFont and QueryTextExtents only).
</para>
<para>
Defense: send Font error.
</para>
<para>
Denial of service: open fonts until the server runs out of memory (OpenFont
only).
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='CloseFont'>
<title>CloseFont</title>
<para>
Destruction: close another client's font.
</para>
<para>
Defense: send Font error.
</para>
</sect1>
<sect1 id='SetFontPath'>
<title>SetFontPath</title>
<para>
Denial of service: change the font path so that other clients cannot
find their fonts.
</para>
<para>
Alteration: change the font path so that other clients get different
fonts than they expected.
</para>
<para>
Defense for both of the above: separate font paths for trusted and
untrusted clients, as described in the Font Requests section.
</para>
<note>
<para>
ISSUE: the printing project considered per-client font paths and concluded
that it was very difficult to do. We should look at this aspect of the print
server design to see if we can reuse the same scheme. We should also try to
reconstruct what was so difficult about this; it doesn't seem that hard on the
surface.
</para>
</note>
</sect1>
<sect1 id='GetFontPath'>
<title>GetFontPath</title>
<para>
Theft: retrieve font path elements that were set by other clients.
</para>
<para>
Use knowledge from font path elements to mount other attacks, e.g.,
attack a font server found in the font path.
</para>
<para>
Defense for both of the above: separate font paths for trusted and
untrusted clients, as described in the Font Requests section.
</para>
</sect1>
<sect1 id='CreatePixmap'>
<title>CreatePixmap</title>
<para>
Theft: resource ID guessing (drawable).
</para>
<para>
Defense: send Drawable error.
</para>
<para>
Denial of service: create pixmaps until the server runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='FreePixmap'>
<title>FreePixmap</title>
<para>
Destruction: destroy another client's pixmap.
</para>
<para>
Defense: send Pixmap error.
</para>
</sect1>
<sect1 id='CreateGC'>
<title>CreateGC</title>
<para>
Theft: resource ID guessing (drawable, tile, stipple, font, clip-mask).
</para>
<para>
Defense: send Drawable, Pixmap, or Font error.
</para>
<para>
Denial of service: create GCs until the server runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='CopyGC'>
<title>CopyGC</title>
<para>
Theft: copy GC values of another client's GC.
</para>
<para>
Alteration: copy GC values to another client's GC.
</para>
<para>
Defense for both of the above: send GC error.
</para>
</sect1>
<sect1 id='ChangeGC_SetDashes_SetClipRectangles'>
<title>ChangeGC, SetDashes, SetClipRectangles</title>
<para>
Alteration: change values of another client's GC.
</para>
<para>
Theft: resource ID guessing (gc, tile, stipple, font, clip-mask)
(last four for ChangeGC only).
</para>
<para>
Defense for both of the above: send GC error.
</para>
</sect1>
<sect1 id='FreeGC'>
<title>FreeGC</title>
<para>
Destruction: destroy another client's GC.
</para>
<para>
Defense: send GC error.
</para>
</sect1>
<sect1 id='Drawing_Requests'>
<title>Drawing Requests</title>
<para>
Specifically, ClearArea, CopyArea, CopyPlane, PolyPoint, PolyLine, PolySegment,
PolyRectangle, PolyArc, FillPoly, PolyFillRectangle, PolyFillArc, PutImage,
PolyText8, PolyText16, ImageText8, and ImageText16.
</para>
<para>
Alteration: draw to another client's drawable.
</para>
<para>
Theft: resource ID guessing: ClearArea - window; CopyArea, CopyPlane -
src-drawable, dst-drawable, gc; all others - drawable, gc.
</para>
<para>
Defense for both of the above: send appropriate error.
</para>
<note>
<para>
ISSUE: The Motif preregister drag protocol requires clients to draw
into windows of other clients for drag-over/under effects.
</para>
</note>
<para>
Spoofing: draw to a window to make it resemble a window of another client.
</para>
<para>
Defense: see
<xref linkend='MapWindow' xrefstyle='select: title'/>
.
</para>
</sect1>
<sect1 id='GetImage'>
<title>GetImage</title>
<para>
Theft: get the image of another client's drawable.
</para>
<para>
Theft: resource ID guessing (drawable).
</para>
<para>
Defense: send Drawable error.
</para>
<para>
Theft: get the image of your own window, which may contain pieces of other
overlapping windows.
</para>
<para>
Defense: censor returned images by blotting out areas that contain data
from trusted windows.
</para>
</sect1>
<sect1 id='CreateColormap'>
<title>CreateColormap</title>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Colormap error.
</para>
<para>
Denial of service: create colormaps with this request until the server
runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='FreeColormap'>
<title>FreeColormap</title>
<para>
Destruction: destroy another client's colormap.
</para>
<para>
Defense: send Colormap error.
</para>
</sect1>
<sect1 id='CopyColormapAndFree'>
<title>CopyColormapAndFree</title>
<para>
Theft: resource ID guessing (src-map).
</para>
<para>
Defense: send Colormap error. However, default colormaps will be allowed.
</para>
<note>
<para>
ISSUE: must untrusted applications be allowed to use standard colormaps?
(Same issue for ListInstalledColormaps, Color Allocation Requests,
FreeColors, StoreColors, StoreNamedColor, QueryColors, and LookupColor.)
</para>
</note>
<para>
Denial of service: create colormaps with this request until the server
runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='InstallColormap_UninstallColormap'>
<title>InstallColormap, UninstallColormap</title>
<para>
Theft: resource ID guessing.
</para>
<para>
Defense: send Colormap error.
</para>
<para>
Denial of service: (un)install any colormap, potentially preventing
windows from displaying correct colors.
</para>
<para>
Defense: treat this request as a no-op. Section 4.1.8 of the ICCCM states
that (un)installing colormaps is the responsibility of the window manager
alone.
</para>
<note>
<para>
ISSUE: the ICCCM also allows clients to do colormap installs if the client
has the pointer grabbed. Do we need to allow that too?
</para>
</note>
</sect1>
<sect1 id='ListInstalledColormaps'>
<title>ListInstalledColormaps</title>
<para>
Theft: resource ID guessing (window).
</para>
<para>
Defense: send Colormap error.
</para>
<para>
Theft: discover the resource ID of another client's colormap from the reply.
</para>
<para>
Defense: remove the returned colormap IDs; only let through default
colormaps and colormaps of untrusted clients.
</para>
</sect1>
<sect1 id='Color_Allocation_Requests'>
<title>Color Allocation Requests</title>
<para>
Specifically, AllocColor, AllocNamedColor, AllocColorCells, and
AllocColorPlanes.
</para>
<para>
Alteration/Denial of service: allocate colors in another client's
colormap. It is denial of service if the owning client's color allocations
fail because there are no cells available. Otherwise it is just alteration.
</para>
<para>
Theft: resource ID guessing (cmap).
</para>
<para>
Defense for both of the above: send Colormap error. However, default
colormaps will be allowed.
</para>
</sect1>
<sect1 id='FreeColors'>
<title>FreeColors</title>
<para>
Theft: resource ID guessing (cmap).
</para>
<para>
Defense: send Colormap error. However, default colormaps will be allowed.
</para>
</sect1>
<sect1 id='StoreColors_StoreNamedColor'>
<title>StoreColors, StoreNamedColor</title>
<para>
Alteration: change the colors in another client's colormap.
</para>
<para>
Theft: resource ID guessing (cmap).
</para>
<para>
Defense for both of the above: send Colormap error. However, default
colormaps will be allowed.
</para>
</sect1>
<sect1 id='QueryColors_LookupColor'>
<title>QueryColors, LookupColor</title>
<para>
Theft: retrieve information about the colors in another client's colormap.
</para>
<para>
Theft: resource ID guessing (cmap).
</para>
<para>
Defense for both of the above: send Colormap error. However, default
colormaps will be allowed.
</para>
</sect1>
<sect1 id='CreateCursor_CreateGlyphCursor'>
<title>CreateCursor, CreateGlyphCursor</title>
<para>
Theft: resource ID guessing (source, mask or source-font, mask-font).
</para>
<para>
Defense: send Pixmap or Font error. However, the default font will be allowed.
</para>
<para>
Denial of service: create cursors until the server runs out of memory.
</para>
<para>
Defense: quotas.
</para>
</sect1>
<sect1 id='FreeCursor'>
<title>FreeCursor</title>
<para>
Destruction: free another client's cursor.
</para>
<para>
Defense: send Cursor error.
</para>
</sect1>
<sect1 id='RecolorCursor'>
<title>RecolorCursor</title>
<para>
Alteration: recolor another client's cursor.
</para>
<para>
Theft: resource ID guessing (cursor).
</para>
<para>
Defense for both of the above: send Cursor error.
</para>
</sect1>
<sect1 id='QueryBestSize'>
<title>QueryBestSize</title>
<para>
Theft: resource ID guessing (drawable).
</para>
<para>
Defense: send Drawable error.
</para>
</sect1>
<sect1 id='ListExtensions_QueryExtension'>
<title>ListExtensions, QueryExtension</title>
<para>
Determine the extensions supported by the server, and use the list to choose
extension-specific attacks to attempt.
</para>
<para>
Defense: extensions will have a way to tell the server whether it is safe
for untrusted clients to use them. These requests will only return information
about extensions that claim to be safe.
</para>
</sect1>
<sect1 id='Keyboard_configuration_requests'>
<title>Keyboard configuration requests</title>
<para>
Specifically, ChangeKeyboardControl, ChangeKeyboardMapping, and
SetModifierMapping.
</para>
<para>
Alteration: change the keyboard parameters that were established by another
client.
</para>
<para>
Denial of service: with ChangeKeyboardControl, disable auto-repeat, key
click, or the bell. With ChangeKeyboardMapping or SetModifierMapping,
change the key mappings so that the keyboard is difficult or impossible to
use.
</para>
<para>
Defense for both of the above: treat these requests as a no-op.
</para>
</sect1>
<sect1 id='Keyboard_query_requests'>
<title>Keyboard query requests</title>
<para>
Specifically, GetKeyboardControl, GetKeyboardMapping, and GetModifierMapping.
</para>
<para>
Theft: get keyboard information that was established by another client.
</para>
<para>
Defense: This is a minor form of theft. We propose to do nothing about this
threat.
</para>
</sect1>
<sect1 id='ChangePointerControl_SetPointerMapping'>
<title>ChangePointerControl, SetPointerMapping</title>
<para>
Alteration: change the pointer parameters that were established by another
client.
</para>
<para>
Denial of service: set the pointer parameters so that the pointer is
difficult or impossible to use.
</para>
<para>
Defense for both of the above: treat these requests as a no-op.
</para>
</sect1>
<sect1 id='GetPointerControl_GetPointerMapping'>
<title>GetPointerControl, GetPointerMapping</title>
<para>
Theft: get pointer parameters that were established by another client.
</para>
<para>
Defense: This is a minor form of theft. We propose to do nothing about this
threat.
</para>
</sect1>
<sect1 id='SetScreenSaver'>
<title>SetScreenSaver</title>
<para>
Alteration: change the screen saver parameters that were established by
another client.
</para>
<para>
Denial of service: set the screen saver parameters so that the screen saver
is always on or always off.
</para>
<para>
Defense for both of the above: treat these requests as a no-op.
</para>
</sect1>
<sect1 id='GetScreenSaver'>
<title>GetScreenSaver</title>
<para>
Theft: get screen saver parameters that were established by another client.
</para>
<para>
Defense: This is a minor form of theft. We propose to do nothing about this
threat.
</para>
</sect1>
<sect1 id='ForceScreenSaver'>
<title>ForceScreenSaver</title>
<para>
Denial of service: repeatedly activate the screen saver so that the user
cannot see the screen as it would look when the screen saver is off.
</para>
<para>
Denial of service: repeatedly reset the screen saver, preventing it from
activating.
</para>
<para>
Defense for both of the above: treat these requests as a no-op.
</para>
</sect1>
<sect1 id='ChangeHost'>
<title>ChangeHost</title>
<para>
Most servers already have some restrictions on which clients can use this
request, so whether the following list applies is implementation dependent.
</para>
<para>
Denial of service: remove a host from the list, preventing clients from
connecting from that host.
</para>
<para>
Add a host to the list. Clients from that host may then launch other
attacks of any type.
</para>
<para>
Defense for both of the above: return Access error.
</para>
</sect1>
<sect1 id='ListHosts'>
<title>ListHosts</title>
<para>
Theft: steal host identities and possibly even user identities that are
allowed to connect.
</para>
<para>
Launch attacks of any type against the stolen host/user identities.
</para>
<para>
Defense for both of the above: return only untrusted hosts.
</para>
</sect1>
<sect1 id='SetAccessControl'>
<title>SetAccessControl</title>
<para>
Most servers already have some restrictions on which clients can use this
request, so whether the following list applies is implementation dependent.
</para>
<para>
Alteration: change the access control value established by some other client.
</para>
<para>
Disable access control, allowing clients to connect who would normally not be
able to connect. Those clients may then launch other attacks of any type.
</para>
<para>
Defense for both of the above: return Access error.
</para>
</sect1>
<sect1 id='SetCloseDownMode'>
<title>SetCloseDownMode</title>
<para>
Denial of service: set the close-down mode to RetainPermanent or
RetainTemporary, then disconnect. The server cannot reuse the
resource-id-base of the disconnected client, or the memory used by the
retained resources, unless another client issues an appropriate KillClient
to cancel the retainment. The server has a limited number of
resource-id-bases, and when they are exhausted, it will be unable to accept
new client connections.
</para>
<para>
Defense: treat this request as a no-op.
</para>
</sect1>
<sect1 id='KillClient'>
<title>KillClient</title>
<para>
Destruction/Denial of service: kill another currently connected client.
</para>
<para>
Destruction: kill a client that has terminated with close-down mode of
RetainTemporary or RetainPermanent, destroying all its retained resources.
</para>
<para>
Destruction: specify AllTemporary as the resource, destroying all resources
of clients that have terminated with close-down mode RetainTemporary.
</para>
<para>
Defense for all of the above: return Value error.
</para>
</sect1>
<sect1 id='Clean_Requests'>
<title>Clean Requests</title>
<para>
Other than denial of service caused by flooding, these requests have no known
security concerns: AllowEvents, UngrabPointer, UngrabButton, UngrabKeyboard,
UngrabKey, UngrabServer, NoOperation, and Bell.
</para>
</sect1>
</chapter>
<chapter id='Events'>
<title>Events</title>
<para>
The only threat posed by events is theft. Selecting for events on another
client's resources is always theft. We restrict further analysis by
assuming that the client only selects for events on its own resources,
then asking whether the events provide information about other clients.
</para>
<sect1 id='KeymapNotify'>
<title>KeymapNotify</title>
<para>
Theft: the state of the keyboard can be seen when the client does not have
the input focus. This is possible because a KeymapNotify is sent to a
window after every EnterNotify even if the window does not have input focus.
</para>
<para>
Defense: zero the returned bit vector if a trusted client currently has the
input focus.
</para>
</sect1>
<sect1 id='Expose'>
<title>Expose</title>
<para>
Theft: discover where other clients' windows overlap your own. For
instance, map a full-screen window, lower it, then raise it. The resulting
exposes tell you where other windows are.
</para>
<para>
Defense: about the only thing you could do here is force backing store to be
used on untrusted windows, but that would probably use too much server
memory. We propose to do nothing about this threat.
</para>
</sect1>
<sect1 id='GraphicsExposure'>
<title>GraphicsExposure</title>
<para>
Theft: discover where other clients' windows overlap your own. For instance,
use CopyArea to copy the entire window's area exactly on top of itself. The
resulting GraphicsExposures tell you where the window was obscured.
</para>
<para>
Defense: see Expose above. We propose to do nothing about this threat.
</para>
</sect1>
<sect1 id='VisibilityNotify'>
<title>VisibilityNotify</title>
<para>
Theft: this event provides crude positional information about other
clients, though the receiver cannot tell which other clients.
</para>
<para>
Defense: The information content of this event is very low. We propose to
do nothing about this threat.
</para>
</sect1>
<sect1 id='ReparentNotify'>
<title>ReparentNotify</title>
<para>
Theft: the parent window may belong to some other client (probably the
window manager).
</para>
<para>
Defense: If the parent window belongs to a trusted client, return the
closest ancestor window that belongs to an untrusted client, or if such a
window does not exist, return the root window for the parent window.
</para>
<note>
<para>
ISSUE: what is the application impact?
</para>
</note>
</sect1>
<sect1 id='ConfigureNotify'>
<title>ConfigureNotify</title>
<para>
Theft: the above-sibling window may belong to some other client.
</para>
<para>
Defense: return None for the above-sibling window if it belongs to a trusted
client.
</para>
<note>
<para>
ISSUE: what is the application impact?
</para>
</note>
</sect1>
<sect1 id='ConfigureRequest'>
<title>ConfigureRequest</title>
<para>
Theft: the sibling window may belong to some other client.
</para>
<para>
Defense: return None for the sibling window if it belongs to a trusted client.
</para>
<para>
ISSUE: what is the application impact?
</para>
</sect1>
<sect1 id='SelectionClear'>
<title>SelectionClear</title>
<para>
Theft: the owner window may belong to some other client.
</para>
<para>
Defense: return None for the owner window if it belongs to a trusted client.
</para>
</sect1>
<sect1 id='SelectionRequest'>
<title>SelectionRequest</title>
<para>
Theft: the requestor window may belong to some other client.
</para>
<para>
Defense: Blocking this event or censoring the window would prevent selection
transfers from untrusted clients to trusted clients from working. We propose
to do nothing in the server about this threat. The security manager may
reduce the exposure of trusted window IDs by becoming the owner of all
selections.
</para>
</sect1>
<sect1 id='MappingNotify'>
<title>MappingNotify</title>
<para>
Theft: discover keyboard, pointer, or modifier mapping information set
by another client.
</para>
<para>
Defense: Any tampering with this event will cause clients to have an
inconsistent view of the keyboard or pointer button configuration, which is
likely to confuse the user. We propose to do nothing about this threat.
</para>
</sect1>
</chapter>
<chapter id='Errors'>
<title>Errors</title>
<para>
There appear to be no threats related to protocol errors.
</para>
</chapter>
<chapter id='Future_Work'>
<title>Future Work</title>
<para>
The next steps are resolve the items marked ISSUE and to decide if the
defenses proposed are reasonable. Discussion on the security@x.org mailing list,
prototyping, and/or starting the implementation should help answer these
questions.
</para>
</chapter>
<chapter id='References'>
<title>References</title>
<para>
Bellcore, "Framework Generic Requirements for X Window System Security,"
Technical Advisory FA-STS-001324, Issue 1, August 1992.
</para>
<para>
Dardailler, Daniel, "Motif Drag And Drop Protocol," unpublished design
notes.
</para>
<para>
Kahn, Brian L., "Safe Use of X WINDOW SYSTEM protocol Across a Firewall",
unpublished draft, The MITRE Corporation, 1995.
</para>
<para>
Rosenthal, David S. H., "LINX - a Less INsecure X server," Sun Microsystems,
29th April 1989.
</para>
<para>
Rosenthal, David and Marks, Stuart W., "Inter-Client Communication Conventions
Manual Version 2.0," ftp://ftp.x.org/pub/R6.1/xc/doc/hardcopy/ICCCM/icccm.PS.Z
</para>
<para>
Scheifler, Robert W., "X Window System Protocol,"
ftp://ftp.x.org/pub/R6.1/xc/doc/hardcopy/XProtocol/proto.PS.Z
</para>
<para>
Treese, G. Winfield and Wolman, Alec, "X Through the Firewall, and Other
Application Relays," Digital Equipment Corporation Cambridge Research Lab,
Technical Report Series, CRL 93/10, May 3, 1993.
</para>
</chapter>
</book>