= Security Considerations = This page exists so that there is a single place to go to learn about the general security properties of Tahoe, as well as about any current known issues that might have security consequences. = Current Known Security Issues in Tahoe = * XSRF / Browser-based Attacks * #127: the URI of a file is embedded in the URL that is used to access it. This URI should not be unintentionally revealed to anyone else, because that would reveal the full contents of the file. There are two current ways this URI can be unintentionally revealed: * If the file is HTML and contains a hyperlink to an external web server, any user who follows that hyperlink may reveal the URI to that web server through the Referrer header. * If the file is HTML and contains active content such as Javascript, that javascript can read the URL (and consequently the secret URI) as it runs. This javascript may then find some clever way to reveal the URI to a third party (such as by changing the src= attribute of an image tag). * We are thinking about ways to close off this leakage of authority while preserving ease of use -- the ticket associated with this issue is ticket #127. In the meantime, a good work-around is to remove all hyperlinks pointing to external servers from any HTML file that you upload to a Tahoe grid, and to not store HTML with embedded javascript, if you want the contents of the file to remain private. Note that no other files or directories are threatened, only the HREF/JS-bearing HTML file. = General Security Properties of Tahoe = '''The rest of this page, below, is not complete.''' However, you can view [source:docs/architecture.txt@1432#L472 the detailed technical explanation] of which this page is eventually intended to be a summary. = The Distributed Filesystem = == Access Control == The Tahoe distributed filesystem is composed of files and directories. === Files === ==== read access ==== Each file has a unique and unguessable identifier, called a "CHK-URI", which may be derived from the file contents. Possession of this identifier is necessary and sufficient to download, reconstruct, decrypt, and verify the integrity of the file. If a person is not given the CHK-URI, then they cannot see the contents of the file. ==== mutation ==== Files in the Tahoe grid are immutable. If you upload a file to the grid, and then change part of it and upload it again, then there are now two files in the grid -- the old one and the new one -- and each has a distinct, unique, CHK-URI. The directory to which the new file was uploaded will only contain a reference to the new file. If no other directories still reference the old file (and if no manual copies of the URI were retained), the old file will be unreachable. A future extension will provide mutable files. For these, a given URI will not necessarily refer to a specific sequence of bytes, but rather to just the most recent contents that were uploaded to that URI. Like dirnode URIs, these URIs will come in read-write and read-only forms, and the file can only be modified by someone who holds a read-write URI. == Traffic Analysis == ''To be filled in.'' Traffic analysis is subtle and powerful. The distributed nature of Tahoe provides even more information to a passive observer than usual. All traffic between tahoe nodes uses transport-level encryption, so an attacker must participate in a Tahoe network to obtain visibility into which shares are being uploaded and downloaded. However, the promiscuous nature of tahoe's Introduction protocol makes this rather easy. In small networks, most server see upload and download requests for all files. In large networks, an attacker who can provide at least 10% of the servers (for 3-of-10 encoding) will get to see upload/download requests for all files. By seeing these requests, the attacker gets to know who is interested in which files, although they cannot determine the contents of those files unless they already have a copy (and convergence is being used). The directory nodes are encrypted, but all of the dirnodes are stored on the same central server (the "vdrive server"). This server is in an excellent position to see who accesses which dirnodes and when, and this information is sufficient to build a dirnode graph that is equivalent to the user's plaintext version. For example, if the server sees a get(dirnode#47, "34af") followed by a get(dirnode#13, "8bb3"), it is safe to assume that dirnode#47 contains dirnode#13 as a subdirectory, and that "34af" is the encrypted form of the subdir's name. This reconstructed graph has file/subdir names which are encrypted but the same length as the real ones. The file URIs are not known, although if a file is uploaded or downloaded shortly after a dirnode is accessed it is easy to relate the two. Again, this points to the identity of the file, but not its contents. However, it makes it fairly easy for the dirnode server to tell, e.g., if a lot of users are all referencing the same file. A future design will include distributed directory nodes (to improve availability and reliability). This will result in the same traffic-analysis exposure as the centralized vdrive server, but makes the traffic visible to even more servers (anyone who controls more than 10% of the servers will be able to see all dirnode requests).