289675 - getter/setter on window.__proto__.__proto__.__proto__ allows XSS attacks

Status: RESOLVED FIXED

(Core :: DOM: Core & HTML, defect)

Description

[shutdown]

User-Agent:       Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1)
Build Identifier: Mozilla/5.0 (Windows; U; Win98; en-US; rv:1.7.6) Gecko/20050317 Firefox/1.0.2

Getter/setter defined on window.__proto__.__proto__.__proto__ is
executed when a global variable is used in scripts,
but is not cleared when a new page is loaded.
Therefore allows XSS attacks.
Testcase available.

Reproducible: Always

Steps to Reproduce:
1.define getter/setter on window.__proto__.__proto__.__proto__ with some name.
2.navigate the browser to the page that uses a global variable with a name 
defined in step 1.
Actual Results:  
Getter/setter defined in step 1 will be evaluated in the context of a new page.

Expected Results:  
Getter/setter defined in step 1 must be cleared when loading a new page.
Or, do not allow web pages to poison window.__proto__.__proto__.__proto__.

Comment 1

[shutdown]

Attachment: testcase

testcase for the bug.

Comment 2

[Daniel Veditz [:dveditz]]

Confirming.

If you set browser.dom.global_scope_pollution.disabled to true (undoing the IE
compat fixes in bug 256932) you can still do this, but using only two __proto__
instead of three.

Comment 3

[Brendan Eich [:brendan]]

That is the last object on the prototype chain, namely Object.prototype.  We
should be resetting that to a new one created for a new Object ctor on new doc
load, of course.

/be

Comment 4

[Johnny Stenback (:jst)]

Attachment: Make context re-initialization work right again.

Not sure when or how this broke, didn't have time to dig into that yet. But the
problem is that XPConnect is unable to find Object.prototype when we're
re-initializing a context (i.e. during page transition), so it ends up using
the old one it had. That's bad, and there was code that printed out a warning
for that, but I hadn't been observant enough to notice that. I turned on that
code for everyone now to hopefully catch this earlier if it happens again.

So this patch makes it so that we can resolve standard classes during context
re-initialization. Pretty straight forward.

Comment 5

[Chase Phillips]

Comment on attachment 180219 [details] [diff] [review]
Make context re-initialization work right again.

a=chase/drivers pending r/sr

Comment 6

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 180219 [details] [diff] [review]
Make context re-initialization work right again.

r=bzbarsky.

Comment 7

[Brendan Eich [:brendan]]

Comment on attachment 180219 [details] [diff] [review]
Make context re-initialization work right again.

I think that NS_WARNING should be an NS_ERROR.	Is that going to turn tinderbox
orange?  If so, we still have a potential security hole.

/be

Comment 8

[Daniel Veditz [:dveditz]]

r=dveditz too, but get rid of the NS_WARNING and uncomment the NS_ERROR

Comment 9

[Daniel Veditz [:dveditz]]

jst checked in to the aviary branch

Comment 10

[Johnny Stenback (:jst)]

Fixed on the 1.7 branch too.

Comment 11

[Mike Connor [:mconnor]]

Attachment: trunk version of what got checked in

Comment 12

[Mike Connor [:mconnor]]

landed on trunk

Comment 13

[Boris Zbarsky [:bzbarsky]]

This caused some pretty weird (though possibly correct) code flow during context
init.  I filed bug 295606 on that.

Comment 14

[timeless]

Comment on attachment 181440 [details] [diff] [review]
trunk version of what got checked in

this patch was applied in cenzic hailstorm 2.6 against mozilla/ -mozilla1.8a5
and js/ -mozilla1.8b1