From: Phil Karlton <karlton@netscape.com>
Date: Sun, 8 Oct 95 01:39:18 PDT
To: cypherpunks@toad.com
Subject: Re: NEW Netscape RNG hole
In-Reply-To: <9510080732.AA14015@anon.penet.fi>
Message-ID: <30778E3D.4CBD@netscape.com>
MIME-Version: 1.0
Content-Type: text/plain


RingZero wrote:
> 
> Did anyone else notice a bug in the new, public Netscape
> RNG code? It appears that on Windows builds, during the
> RNG seeding, the function that hashes in file contents
> (EnumSystemFiles) doesn't close a file handle (lFileHandle).

I think you mean lFindHandle. I'm not a windows programmer, so I have
no idea if the enumerator needs to be cleaned up, but I will forward
your message to the appropriate folks here.

> This doesn't hurt too badly on the client, but on a server,
> leaking these resources is deadly.

Not really. SEC_SystemInfoForRNG is called ONCE for each execution.
This is true for servers and for clients. Did you rig up a test
bed that called SEC_SystemInfoForRNG in a loop? In addition, if
the server cannot open any more files, it will have difficulty
delivering any information (compromised or not) to any clients.

The clients use the SEC_GetNoise data each time they go idle to
re-initialize the PRNG seed. The servers also re-initialize the
PRNG seed periodically. That function does not use any file
handles.

> I ran some experiments. It took a few thousand calls before
> these open file handles forced not only the file content function
> to fail, but also made OTHER calls quietly fail. With these calls
> quietly failing, the RNG is significantly weakened. In my tests
> on Windows NT, ALL of the following RNG functions failed:
> 
> * GetComputerName
> * GetVolumeInformation
>         volume Name,
>         volume Serial Number,
>         Maximum Filename Length
>         Filesystem Flags
>         Filesystem Name
> * GetDiskFreeSpace
>         SectorsPerCluster
>         BytesPerSector
>         Free Clusters
>         Total Clusters
> * subroutine for the inclusion of system files, both number of them & 
contents
>         ReadSystemFiles()
> * subroutine used for other history file accesses
>         SEC_FileForRNG(*filename)

SEC_FileForRNG does close every file it opens. As I said before,
I don't know what the rules are for closing the enumerators.
 
> How did this get past Netscape testers?

The tested servers worked? :-)

> Does anyone
> know if this was fixed before Netscape shipped?

The code that was made available for download was the same that
was in the shipped updates.

> Does
> it rate a shirt, or does this mean  Jeff W. gets to shave his
> head? I seem to remember him promising to shave it if we
> could show a significant weakness in the new RNG code,
> and since this does (IMO)...

That was me (Phil Karlton) and not Jeff who put his locks (pun intended)
at risk. You haven't demonstrated to me that there is significant
weakness. I'll keep my hair for another day.

I don't think you are going to get a shirt. It's hard to mail one
to an anonymous person. On the other hand, they are not mine to
give or withhold.

PK
--
Philip L. Karlton			karlton@netscape.com
Principal Curmudgeon			http://www.netscape.com/people/karlton
Netscape Communications Corporation