- From: John Boyer <jboyer@uwi.com>
- Date: Wed, 17 Nov 1999 16:59:11 -0800
- To: "Greg Whitehead" <gwhitehead@signio.com>, "DSig Group" <w3c-ietf-xmldsig@w3.org>
Hi Greg, Sorry, but it's not a "clumsy" solution to suggest that A) transforms can go on SignedInfo but B) there must be some restrictions. I suggested transforms on SignedInfo not because I wanted to be crucified but because I wanted to be very specific about which transforms can be omitted. Read Mark Bartel's recent post on the dangers inherent in allowing some ObjectReference transforms to be omitted. The SignedInfo transforms would serve as a precise description of what got omitted. If too much of ObjectReference gets omitted, then the digital signature is thrown out as worthless. But if the SignedInfo transforms omit only transforms of type base-64, then document closure and overall security are preserved. The other reason I suggested doing this by transforms is that it seemed unclean from an engineering standpoint to create a second method to omit Location and certain transforms when we already have a method. Finally, it may be quite possible to slightly tweak the markup design so that these "clumsy" security exceptions don't even exist. That's why it's not a good idea to use such characterizations too quickly. John Boyer Software Development Manager UWI.Com -- The Internet Forms Company Hi John, The question isn't so much whether transforms are signed or not, but what they're allowed to do. I have to admit a certain ignorance of the capabilities of the transforms proposed for inclusion in the initial version of the spec, but I'm assuming that once Pandora's box is opened anything is possible. The bottom line is that any transform over SignedInfo with the ability to modify itself (along with some other critical piece of information like SignatureMethod or the DigestValue or Transforms in an ObjectReference) is open to abuse. Given all of the restrictions that must be placed on transforms over SignedInfo in the name of security, it seems like a clumsy solution to the original problem. I'm not discounting the original problem -- I just prefer a solution that passes the object location problem off to the application. Even in the case of URLs pointing to resources that haven't moved, an application verifying the signature on an object probably has other uses for the object itself (and should probably verify the signature against the version of the object that it fetches, rather than a version that the signature verification core fetches independently). -Greg -----Original Message----- From: John Boyer [mailto:jboyer@uwi.com] Sent: Wednesday, November 17, 1999 2:30 PM To: Greg Whitehead; DSig Group Subject: RE: Omitting Location and Transforms from SignedInfo Hi Greg, I apologize for the delay; I've had a lot of work to overcome recently. I almost agree with your entire analysis. Basically, in order to hijack the signature, it seems that you would have to create a SignedInfo transform that omits itself as well as does the interesting things you described. By contrast, I rather assumed that we would create a bottom turtle (to borrow one of Rohit's favorite phrases, and also to follow the irony that the problem comes from self-reference, a favorite of Russell's). By bottom turtle I mean that SignedInfo should include its Transforms in its text to be hashed even if the Transforms say to omit themselves. I faced this problem two years ago in XFDL, and the solution was to disregard the filters if they tried to filter themselves out of the message to be digested. Actually, we do have a larger problem now because of that awful XSLT transform. I only want the transforms for the xpath transform because I only want keep/omit abilities on the given markup, not the ability to turn it into Mohawk. Anyway, my point is that I agree with your analysis as given, but the problem is the result of differing assumptions. I assume that the SignedInfo transforms are automatically covered by the SignatureValue and hence are not modifiable. Under that assumption, the security problem vanishes before it gets started. John Boyer Software Development Manager UWI.Com -- The Internet Forms Company <Greg/> Forget that, however, and consider this! You sign object x without specifying any transform over SignedInfo (the signature includes all SignedInfo elements): <Signature> <SignedInfo> ... <ObjectReference location="x"> ... <DigestValue>xxx</DigestValue> </ObjectReference> </SignedInfo> ... </Signature> I modify your signature to refer to object y and I introduce a transform that undoes this change (restores the original contents of the SignedInfo): <Signature> <SignedInfo> <Transforms> <Transform> replace SignedInfo with original contents </Transform> </Transforms> ... <ObjectReference location="y"> ... <DigestValue>yyy</DigestValue> </ObjectReference> </SignedInfo> As I understand the proposed verification logic, I have now hijacked your signature. Specifically, we will first go find object y and verify that its digest is yyy. We will then apply the transformation to SignedInfo, which will restore it to what you originally signed. We will then verify your signature of the transformed SignedInfo, which will succeed (implying that you signed y). You can't solve this problem by suggesting that verification of digest values should take place after the transformations are applied, since you are proposing to drop information like location. Getting back to the need to drop location at all, your objection to application specific object location logic is that it can't be cleanly integrated with the core signature verification mechanism. I don't buy that. It's a lot cleaner to use a callback to locate objects than to rely on re-writing the ObjectReferences within the signature. -Greg -----Original Message----- From: John Boyer [mailto:jboyer@uwi.com] Sent: Friday, November 12, 1999 10:05 AM To: Greg Whitehead; DSig Group Subject: RE: Omitting Location and Transforms from SignedInfo Hi Greg, This is not the only concrete example. If you go back to the letter I sent (or indeed my IETF presentation conclusion) you will find that it is not only possible for transforms to reduce security but also to obliterate it entirely. Rather than omitting the SignatureMethod, simply omit the DigestValue from ObjectReferences, which disconnects the signer from the signed. Still, no application is going to create such signatures precisely because they will make such newsworthy fodder for the Bruce Schneier types of the world. Essentially, signing SignedInfo is an implied ObjectReference. Whatever fears we have about omission from SignedInfo are precisely the same fears for omission from referenced objects. It is a dangerous but necessary feature (and yes I can explain why (again) in painstaking detail for those who require it, or you could just come to heckle my presentation of the material at RSA2000). John Boyer Software Development Manager UWI.Com -- The Internet Forms Company I just thought of a concrete example: SignatureMethod is included in SignedInfo to protect against a downgrade attack, should one of the currently approved signature methods be broken. Allowing arbitrary transformation of SignedInfo could potentially defeat this protection by allowing an attacker to introduce a transformation that substitutes in a broken SignatureMethod along with a reference to a modified object and other changes (exploiting the broken signature method to produce the original SignatureValue over the modified SignedInfo). -Greg
Received on Wednesday, 17 November 1999 20:00:10 UTC