By John Gruber
[Update 24 May: If you just want to know the steps I recommend to close the various URI-related vulnerabilities, see “An Ounce of Prevention”.]
[Update 23 May, 11:45 EST: This article has been revised significantly. Ordinarily, I do not make significant changes to an article in-place after publishing, but in this case, I believe accuracy is paramount. I’ve saved a copy of the original article (in Markdown format), if you’re interested.]
Apple’s description for the Panther version:
This update delivers a number of security enhancements and is recommended for all Macintosh users. This update includes the following components:
The description for the Jaguar version:
Security Update 2004-05-24 delivers a number of security enhancements and is recommended for all Macintosh users. This update includes the following components:
Indeed, these are the only updated components. (You can see for
yourself by downloading the updates. The
.pkg updaters are
packages, which are just folders that the Finder treats as a single
file. Inside the package is a “Contents” folder, inside which is a
file named “Archive.bom”. (“bom” stands for “bill of materials”.)
You can use the
lsbom command-line tool to list all of the updated
files specified in a
The updated version of Help Viewer for Panther is version 2.0.6. The new version of Help Viewer fixes at least one specific security exploit: Help Viewer no longer executes scripts when it receives a ‘help:runscript’ URI which originates from any application other than Help Viewer itself.
Thus, after installing this security update, it is now safe to assign the ‘help’ URI protocol to Help Viewer.
When I first wrote about this ‘help’ URI security exploit, I mentioned that I wasn’t aware of any help books which actually took advantage of this feature. That was sheer laziness on my part, and for that I apologize. In fact, as a simple BBEdit multi-file search for “help:runscript” shows, it’s a commonly used feature in many of Apple’s own help books.
Most frequently, it’s used to create a link that opens an application or a particular System Prefs panel. For example, each of the Mac Help pages relating to sound — choosing an alert sound, setting the volume, setting input sources, etc. — has a link at the bottom that says, “Open Sound preferences for me”. If you click this link, the System Prefs application launches, and displays the Sound panel.
Help books are just HTML files. The source code for such a link looks like this (note that I’ve added line breaks for readability; the actual href attribute should be one long line):
<a href="help:runscript= MacHelp.help/Contents/Resources/English.lproj/shrd/OpnApp.scpt%20 string='System:Library:PreferencePanes:Sound.prefPane'"> Open Sound preferences for me</a>
Because Help Viewer is the default handler for the ‘help’ URI protocol, when any other application attempts to resolve such a URI, it gets passed to Help Viewer. Prior to the version released in this security update, you could pass a ‘help:runscript’ URI from any source, and Help Viewer would execute the script specified in the href attribute.
Hence the security problem — the ‘disk’ and ‘disks’ URI protocols, ostensibly unrelated to ‘help’ — could be used by a malfeasant to silently mount a disk image on your computer. Then, a ‘help:runscript’ URI could be used to launch apps on the mounted disk image. This could happen simply by loading a web page; with a couple of frame and meta refresh tags, the disk image could mount and the script could execute without your having to click on anything.
The stop-gap measure I recommended prior to the release of the security update was to disable the ‘help’ URI protocol, using RCDefaultApp (or More Internet or MisFox; any of which would do the job). Doing so plugged the security hole, but also prevented any ‘help’ URIs originating outside Help Viewer from working. For example, a ‘help:search’ URI will perform a search in Help Viewer; e.g. to search for “sound volume”:
<a href="help:search='sound%20volume'%20bookID='Mac%20Help"> Search Help Viewer for information about 'sound volume'.</a>
Disabling ‘help’ disabled all help URIs; but the only dangerous ones were ‘help:runscript’ URIs. This is exactly what the new version of Help Viewer limits. You can try it for yourself. Here are the two links shown above:
The first — a ‘help:runscript’ URI — no longer works when initiated from a web browser. Prior to the security update, it did. (When passed a ‘help:runscript’ URI from another app, the updated Help Viewer now logs a message to your console: “Help Viewer help://runscript called by another application!”.)
The second — a ‘help:search’ URI — still works from any application, if you have Help Viewer specified as your default handler for ‘help’. This is a good thing; Apple has prevented precisely what was dangerous, and nothing else.
It’s also worth noting that even when you disable the ‘help’ protocol, or point it at a dummy application such as Chess, Help Viewer still handles all ‘help’ URIs which originate from within Help Viewer itself. This makes sense, in the same way that web browsers like Safari, Camino, and IE always handle ‘http’ URIs on their own, no matter if they’re specified as your default web browser.
Thus, if you’re paranoid, you can leave the ‘help’ protocol disabled, and ‘help:runscript’ URIs will still work for you from within Help Viewer. For what it’s worth, I have restored Help Viewer as my default handler for ‘help’.
The updated Help Viewer for Jaguar has been changed similarly to the version for Panther: scripts are no longer executed when ‘help:script’ URIs originate from an app other than Help Viewer itself.
I have confirmed that the updated version of Terminal for Jaguar fixes the ‘telnet’ vulnerability. I do not know why this was fixed for Jaguar, but not for Panther.
Security Update 2004-05-24 for Panther contains nothing in response to the ‘telnet’ URI vulnerability I wrote about Friday. My advice remains that you should use RCDefaultApp to disable the ‘telnet’ protocol. (Unless you’re running Jaguar, in which case the updated version of Terminal removes the vulnerability.)
As for ‘disk’ and ‘disks’, you should leave these protocols disabled as well. The specific security problem that led to this discussion has been closed with the updated version of Help Viewer, but it’s a terrible idea to allow disk images to mounted automatically in response to a URI from a remote source, and it can be exploited for harm. E.g., see the first example exploit linked from Unsanity’s Paranoid Android Whitepaper.
But first, what is Paranoid Android? Unsanity describes it thusly:
A vulnerability in Apple’s Mac OS X results in a potential situation in which a malicious person could execute arbitrary commands on your machine, such as deleting your home directory, or doing other harmful actions. This vulnerability involves the use of URL “schemes”. These are the part of a web address that specifies what program should be used to handle the address.
Paranoid Android can protect you from this potential vulnerability until Apple makes an official fix available. It does this by watching the URL schemes that are requested and delaying them until you’ve had a chance to say whether you’d like to proceed or not. If you know that the url that’s being loaded is legit, go ahead, but if it looks suspicious, Paranoid Android gives you an opportunity to cancel it.
But that only describes what Paranoid Android does, not what it is or how it works. What it is is an Application Enhancer module, a.k.a. a “haxie”. How it works, loosely, is that it injects code and/or data into every running application — while haxies do not modify the applications on disk, they do their thing by violating the boundaries of protected memory.
Thus, Paranoid Android is a haxie that injects code into every application, which code apparently intercepts calls to Launch Services which are attempting to resolve URIs. It then presents a confirmation dialog for any URI using a scheme that isn’t known to be safe.
Thus, while Paranoid Android may well solve the problem of malicious use of URI protocols and their default handlers as registered in Launch Services, it may also introduce problems of its own.
The Paranoid Android Whitepaper, written by the developer of Paranoid Android, Jason Harris, documents and links to two “benign sample exploits”. Both are legitimate threats. The first uses the ‘disk’ protocol, and works like this:
A Mac OS X user visits a web page. (It doesn’t matter which web browser is being used.)
Using frames and meta refresh tags, the web page sends a ‘disk’ URI, which, with default settings, causes a disk image to be automatically mounted on the user’s system.
This disk image contains an application bundle, the Info.plist file of which specifies that this app wants to register a custom URI scheme. For example, in Unsanity’s sample exploit, the pertinent snippet of Info.plist is this:
<key>CFBundleURLTypes</key> <array> <dict> <key>CFBundleURLName</key> <string>Malware URL</string> <key>CFBundleURLSchemes</key> <array> <string>malware</string> </array> </dict> </array>
which registers the application as the default handler for the ‘malware’ URI scheme. This could be any arbitrary string — the point is that an application can register for its own custom scheme.
Upon displaying this application in a Finder window, the Finder automatically uses Launch Services to register the application for the URI scheme it has asked to handle. (As far as I can tell, the Finder only does this if the app has been displayed — merely mounting the image doesn’t register the URI scheme in my testing.)
After waiting a few seconds for the disk image to mount, the remote server sends a URI using the protocol the application just registered for. (E.g. in the case of Unsanity’s example, a “malware:unused” URI.) The user’s web browser passes the URI to Launch Services for resolution, and Launch Services launches the application.
At this point, the user’s goose is potentially cooked — once the application is launched, it has free reign. The only action the user took was to load a web page; the rest happened automatically.
Thus, Unsanity’s whitepaper describes a legitimate threat. However, I do not agree with their advice as to how to deal with this threat:
Because this sample exploit registers its own URL scheme, none of the methods people had been using involving disabling certain scripts, moving Help.app or changing the ‘help’ URL scheme would protect against it. At this time, only Paranoid Android provides protection from it.
While indeed this has nothing to do with the ‘help’ scheme, it is simply not true that “only Paranoid Android provides protection from it.” If you have disabled the ‘disk’ and ‘disks’ protocols — and turned off the “Open ‘safe’ files” preference in your web browser — this exploit will fail, because the disk image will not be mounted.
Unsanity’s second sample exploit, added today, uses an ‘ftp’ URI pointing to a directory on unsanity.com. With default settings on Panther, ‘ftp’ URIs are handled by the Finder, which will in turn mount the specified directory in the ‘ftp’ URI as though it were a network server.
Once this happens, and the Finder displays the “OSXMalware” application in that folder, it will once again register the custom URI scheme specified in the app’s Info.plist file. (In this case, ‘guardian452’, which is a different protocol than the ‘malware’ scheme used in Unsanity’s ‘disk’ example.)
My testing indicates this is also a legitimate threat. The app won’t get registered with Launch Services until the Finder displays it, but I can make this happen automatically by invoking Unsanity’s exploit with no Finder windows open; upon mounting the FTP directory as a server volume, the Finder will create a new Finder window displaying the directory’s contents.
When the web server subsequently sends a URI using the newly-registered scheme, Launch Services launches the remote application from the FTP server. Boom.
(Earlier revisions of this article indicated otherwise; at the time this article was originally published, Unsanity’s ‘ftp’ example exploit failed because the OSXMalware application didn’t have executable permissions. They’ve since fixed the permissions, and their example now works.)
Choosing any application other than the Finder as the default handler for ‘ftp’ URIs closes this hole in the same way that disabling the ‘disk’ and ‘disks’ protocols closes the first one. Regular FTP apps do not mount remote FTP directories as though they were server volumes.
This is a generalized and serious vulnerability in Mac OS X:
Remote web server causes a volume to be mounted in the file system, and the contents of the volume are displayed by the Finder.
The now-mounted remote volume contains a malicious application that contains an Info.plist file that asks to register a custom URI scheme with Launch Services.
Upon displaying the malicious application, the Finder registers the new URI scheme, as per the app’s Info.plist file.
After waiting a few seconds for steps 1-3 to occur, the remote web server sends another URI, using the newly-registered scheme.
Launch Services will launch the remote application.
The key to prevention is to nip it at step #1, and prevent remote servers from automatically mounting volumes in your file system.
To protect your Mac, you should definitely disable the following URI protocols, using RCDefaultApp:
You should also assign the ‘ftp’ protocol to any application other than the Finder. (Or disable it, but I think that’s overkill.)
‘afp’ is the scheme for AppleShare servers; I have confirmed that an ‘afp’ URI in the following form will allow a remote AppleShare volume to be mounted automatically and silently:
Note that disabling the ‘afp’ protocol using RCDefaultApp will not prevent you from connecting to AppleShare servers manually. You can still connect to AppleShare servers using the Finder’s Connect to Server command, or using the Network dingus in a Finder window sidebar. Disabling the ‘afp’ protocol merely prevents ‘afp’ URIs from being passed to the Finder from other applications — e.g. a web browser.
You must also make sure your web browser and ‘ftp’ handler do not automatically expand or process quote-unquote “safe” files.
I cannot recommend the use of Paranoid Android. As far as I have determined (and admittedly, that’s an important disclaimer), every vulnerability handled by Paranoid Android can also be solved by changing or disabling Mac OS X’s default URI handlers using RCDefaultApp.
RCDefaultApp merely modifies your Launch Services settings, and should therefore conflict with nothing. Paranoid Android uses completely unsupported mechanisms to inject code into every running application, and therefore has the potential to conflict with anything.
That said, I have no knowledge of any problems or conflicts that Paranoid Android does cause. I believe Jason Harris and Unsanity have nothing but the best of intentions, and their example exploits linked from the Paranoid Android Whitepaper were a significant source of information for this article.
But my point is this: why run a haxie if you can close these vulnerabilities without one?
If there is a demonstrable exploit that cannot be disabled by supported means, I’d like to see it.