572129 - (CVE-2011-2984) Arbitrary code execution via browser's tab element that was dropped on the content area

Status: RESOLVED FIXED

(Core :: Security, defect)

Description

[moz_bug_r_a4]

If a user drag a browser's tab element and drop it on the content area, a
content-registered drop event listener can access the tab element and its
properties including privileged functions.

Only drop event listeners can access privileged objects without the security
check.
http://hg.mozilla.org/mozilla-central/file/2348611d0230/content/events/src/nsDOMDataTransfer.cpp#l455

Comment 1

[moz_bug_r_a4]

Attachment: testcase - Arbitrary code execution

This works on Firefox trunk, 3.6.* and 3.5.*.

(SeaMonkey does not store a tab element in a dataTransfer when a tab is
dragged.)

Comment 2

[Blake Kaplan (:mrbkap) (inactive)]

There are two bugs here:
  * XPCConvert::NativeInterface2JSObject does not create COWs automatically. IMO we should fix this in another bug.
  * nsDOMDataTransfer::MozGetDataAt doesn't do anything to censor chrome objects from content getting at them. While having COWs would fix this bug, IMO it would be better to avoid exposing chrome objects to content at all from this API. Does that sound reasonable to people?

Comment 3

[Blake Kaplan (:mrbkap) (inactive)]

(and the first reason there means that bug 572130 can't save us here.)

Comment 4

[Olli Pettay [:smaug][[e-mail]]]

Sounds very reasonable, and that is what I was thinking too; we shouldn't 
expose chrome objects to content. I'm not quite sure how to filter out
chrome objects.

The first reason is the most important, though.

Comment 5

[Olli Pettay [:smaug][[e-mail]]]

Ah, we have principal for the TransferItems. We could use those.
And content shouldn't even know if there are items which it can't access.

Comment 6

[Olli Pettay [:smaug][[e-mail]]]

Though, perhaps there are cases like string data drop when content should
be able to access the data even if the principal is system principal.
...need to investigate.

Comment 7

[Neil Deakin]

>   * nsDOMDataTransfer::MozGetDataAt doesn't do anything to censor chrome
> objects from content getting at them. While having COWs would fix this bug, IMO
> it would be better to avoid exposing chrome objects to content at all from this
> API. Does that sound reasonable to people?

What are 'chrome objects'?

The only thing that should be prevented is dom objects that aren't accessible from the page the drop listener is in. I assumed that was already happening. Is that not the case?

Comment 8

[Blake Kaplan (:mrbkap) (inactive)]

(In reply to comment #7)
> What are 'chrome objects'?

Objects that are implicitly associated with a chrome "context." In the testcase here, the chrome object is the XUL tab element that is being dropped.

> The only thing that should be prevented is dom objects that aren't accessible
> from the page the drop listener is in. I assumed that was already happening. Is
> that not the case?

It is not, unfortunately. Bug 572647 will help, though.

Comment 9

[Olli Pettay [:smaug][[e-mail]]]

(In reply to comment #8)
> (In reply to comment #7)
> > What are 'chrome objects'?
> 
> Objects that are implicitly associated with a chrome "context."
But what objects are associated with chrome context?
All nsPIDOMEventTargets, sure, but what else?
All the JS implemented (implemented in chrome JS) objects?

Comment 10

[Olli Pettay [:smaug][[e-mail]]]

So it is not clear to me what to do with this.
Bug 572647 would fix the security problem.

Comment 11

[LegNeato]

Is it more likely to get a fix for bug 572647 than this bug? If so we'll track the other one for the branches

Comment 12

[LegNeato]

We're punting this out of .7/.11 for now, please let me know if it should be pulled back (or if bug 572647 should be on the branches instead)

Comment 13

[Olli Pettay [:smaug][[e-mail]]]

So the testcase doesn't seem to work on trunk, but does on branches.

Comment 14

[Damon Sicore (:damons)]

Marcia, can you confirm comment 13?  If it's no longer repro on trunk, let's remove blocking status.

Comment 15

[moz_bug_r_a4]

I can reproduce the testcase on trunk fx-4.0b7pre-2010-10-06-04.

Comment 16

[Marcia Knous [:marcia]]

Yes, I can repro as well using  Mozilla/5.0 (Windows NT 5.1; rv:2.0b7pre) Gecko/20101006 Firefox/4.0b7pre.

Comment 17

[Jonas Sicking (:sicking) No longer reading bugmail consistently]

If compartments fixes this then it should be fixed on tracemonkey by now. Need to verify that before we mark this one FIXED.

However we still need a different fix for branches.

Comment 18

[Blake Kaplan (:mrbkap) (inactive)]

Compartments might not have fixed it due to default-unsafe exposedProps...

Comment 19

[moz_bug_r_a4]

The testcase works on tracemonkey 2010-10-13-04.

Comment 20

[Jonas Sicking (:sicking) No longer reading bugmail consistently]

What do you mean by "works"? Do you mean that you can still successfully get elevated privileges? Or do you mean that the browser successfully stops you?

Comment 21

[moz_bug_r_a4]

The testcase can still successfully get elevated privileges.

Comment 22

[Johnny Stenback (:jst)]

This will actually be fixed by mrbkap's fix in bug 611485 (fixed in tracemonkey). Bug 572647 has been fixed for some time now, but it, as filed, wasn't quite enough to fix this issue as well.

Comment 23

[moz_bug_r_a4]

The old testcase was fixed.  But, this bug's problem is not fixed.  Content
code can still access a tab element and abuse its functions.

Comment 24

[moz_bug_r_a4]

Attachment: testcase 2 - Arbitrary code execution

Comment 25

[Olli Pettay [:smaug][[e-mail]]]

I can certainly reproduce the problem using testcase 2.

Comment 26

[Johnny Stenback (:jst)]

Blake, can we do a spot fix for this issue until COW's are default safe?

Comment 27

[Blake Kaplan (:mrbkap) (inactive)]

Not for FF4. The best I could come up with would be to try to flip "default safe" based on how content code got its hands on the chrome object in question.

Comment 28

[Blake Kaplan (:mrbkap) (inactive)]

Although, I guess we could try to avoid handing out chrome-compartment stuff to content from the drag/drop API... Not sure how hard that would be though or if it'd have other effects.

Comment 29

[Blake Kaplan (:mrbkap) (inactive)]

Attachment: Possible fix

Neil, do you remember why this check was here? It seems to date all the way back to the original implementation.

Comment 30

[Neil Deakin]

Comment on attachment 512386 [details] [diff] [review]
Possible fix

The line is there so that you can drop something from one source to another. Drag and drop wouldn't be particularly useful without this.

Comment 31

[Olli Pettay [:smaug][[e-mail]]]

Should we have some way to mark chrome only DataTransfer data items?
Or perhaps other way, if chrome wants to allow content to access the data,
it should mark the data somehow.
This all at least with "uncommon" data transfer formats.

Comment 32

[Johnny Stenback (:jst)]

Blake said he's got a patch in progress for limiting where chrome objects are handed out from drag n drop code. Blake, please provide an ETA for the patch.

Comment 33

[Blake Kaplan (:mrbkap) (inactive)]

Attachment: Possible fix v2

So, the bug here is really that chrome objects don't behave according to the same origin model. This patch preserves the original behavior for content objects, but denies access to chrome objects that have been dropped.

A second thought: would it make sense in the future to return null instead of throwing here? This seems like a case where the API really means "give me what I asked for if I have access to it, otherwise I don't want to know."

Comment 34

[Neil Deakin]

Comment on attachment 514366 [details] [diff] [review]
Possible fix v2

But this patch has the same problem.

1. Open a page which listens for the drop event:
data:text/html,<body ondragover="return false" ondrop="event.preventDefault(); alert(event.dataTransfer.getData('text/plain'));">Drop here</body>

2. Drag text from the url field or another application onto it.
Actual: Security Exception
Expected: An alert with the dropped text.

Note that the mPrincipal isn't the principal of the thing being dragged, it's the principal of the caller that put it in the dataTransfer (although a native caller could theoretically use any principal).

You actually want to be checking the thing being dragged (stored in mData) to check if it can be accessed. For this specific bug about dropping tabs, that thing will be a <tab> element from chrome.

Comment 35

[Blake Kaplan (:mrbkap) (inactive)]

I'm not really interested in playing whack-a-mole like that. I give up.

Comment 36

[Johnathan Nightingale [:johnath]]

(In reply to comment #35)
> I'm not really interested in playing whack-a-mole like that. I give up.

Where does that put us then, gents?

Comment 37

[Olli Pettay [:smaug][[e-mail]]]

So if we end up blacklisting, do we know what all should be blacklisted.

Or, hmm, could we whilelist things which can be dropped using
system principal?

Comment 38

[Olli Pettay [:smaug][[e-mail]]]

What would happen if the drag started in the content page and
script there sets some data using application/x-moz-tabbrowser-tab?

Comment 39

[Olli Pettay [:smaug][[e-mail]]]

Enn, Gavin, do you happen to know what all drag formats chrome uses?

Comment 40

[Olli Pettay [:smaug][[e-mail]]]

Attachment: Be very strict when handling common DOM objects as transfer data

This is not yet tested properly, but at least dropping files from
desktop to content works, d&d tabs works, d&d text from location bar to content page works etc.

Comment 41

[Olli Pettay [:smaug][[e-mail]]]

So the patch prevents dragging EventTarget objects from content to chrome
and vise versa. Are there cases when we would want to enable that?

Comment 42

[Olli Pettay [:smaug][[e-mail]]]

Comment on attachment 514555 [details] [diff] [review]
Be very strict when handling common DOM objects as transfer data

I really need feedback from those who know Firefox UI code.

Comment 43

[Neil Deakin]

Why is the mExplicitlySet flag needed?

> So the patch prevents dragging EventTarget objects from content to chrome
> and vise versa. Are there cases when we would want to enable that?

I don't think there is currently.

Don't we want to prevent access if the caller cannot access the node no matter whether it is chrome or not? For instance, a drag from one page to another?

Comment 44

[Olli Pettay [:smaug][[e-mail]]]

(In reply to comment #43)
> Don't we want to prevent access if the caller cannot access the node no matter
> whether it is chrome or not? For instance, a drag from one page to another?
Ah, probably. So not only currentIsSystem == dataIsSystem check, but also
principal subsumes check.

and ok, mExplicitlySet isn't probably needed after all...

Comment 45

[Olli Pettay [:smaug][[e-mail]]]

Attachment: v2

Comment 46

[Johnny Stenback (:jst)]

Over to Olli who's working on the alternate fix here!

Comment 47

[Olli Pettay [:smaug][[e-mail]]]

Attachment: a bit simpler per bz's comments on IRC

Comment 48

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 514819 [details] [diff] [review]
a bit simpler per bz's comments on IRC

r=me

Comment 49

[Olli Pettay [:smaug][[e-mail]]]

http://hg.mozilla.org/mozilla-central/rev/bc901af701c9

Comment 50

[Daniel Veditz [:dveditz]]

Comment on attachment 514819 [details] [diff] [review]
a bit simpler per bz's comments on IRC

Approved for 1.9.2.15 and 1.9.1.18, a=dveditz for release-drivers

Comment 51

[Daniel Veditz [:dveditz]]

Comment on attachment 514819 [details] [diff] [review]
a bit simpler per bz's comments on IRC

Not sure why this didn't make it, but moving approvals forward since it missed.

Comment 52

[Daniel Veditz [:dveditz]]

Comment on attachment 514819 [details] [diff] [review]
a bit simpler per bz's comments on IRC

Doesn't compile on 1.9.2, will need a new patch.

  error: ‘class nsIScriptContext’ has no member named ‘GetObjectPrincipal’

Comment 53

[Daniel Veditz [:dveditz]]

Olli: still need that new patch.

Comment 54

[Olli Pettay [:smaug][[e-mail]]]

Attachment: GetObjectPrincipal method from trunk inline

Comment 55

[Daniel Veditz [:dveditz]]

Comment on attachment 549940 [details] [diff] [review]
GetObjectPrincipal method from trunk inline

Approved for 1.9.2.20, a=dveditz

Comment 56

[Olli Pettay [:smaug][[e-mail]]]

http://hg.mozilla.org/releases/mozilla-1.9.2/rev/ab26f1cefe55