oss-sec mailing list archives
Re: Security risk of server side text editing in general and vim.tiny specifically
From: Solar Designer <solar () openwall com>
Date: Fri, 3 Nov 2017 15:18:49 +0100
On Fri, Nov 03, 2017 at 12:57:25PM +0000, Fiedler Roman wrote:
I want to lessen possible legal issues
Oh. I wouldn't have guessed.
The bug may be in the documentation/specification: in my opinion, documentation of good, security aware software should a) implement things considering security bordercases (vim.tiny reporting, that a file was replaced or symlink encountered, proceed?)
Those special cases you list are just a tip of the iceberg.
or b) state, they are not made for that purpose. Even when such statements are redundant for many different tools, they give users at least the chance to learn, that an operation is dangerous and may link to additional information, e.g. the link you provided below on secure root file access. Why has each plastic bag of a new consumer device printed "There is a risk that children pull them over their head and suffocate." for safety reasons,
My guess is mostly for legal reasons, although safety was also involved at some point.
but in software development, we assume, everybody knows and do not include such warnings at least in the footer of man pages?
I don't assume everybody knows. On the contrary, I know that most people don't know, nor do they want to know. When I tell, or ask my fellow sysadmins to follow safer practices, they just get annoyed, in part because the safer practices are too complicated, too brittle, and sometimes also not perfectly safe. But do we really need to include this in every man page? I wish there were a better place.
Editing of non-root files by root should be safe (or be made safe by making changes to the editors where necessary) only in the rare special case when those files are located in a trusted directory. For example, editing as root /var/run/foo owned by user foo should be safe as long as /, /var, and /var/run are owned by root, but editing as root /home/foo/foo or /tmp/foo is unsafe and is likely to stay so.I would need to check that on vim.tiny. As stat-ing, getxattr, renaming, chmod, ... are not atomic, I am not sure if vim.tiny as example would fulfil your expectations. But before that: why do you expect the software to behave like that, when it is not stated anywhere?
What I said in the paragraph quoted above is that I expect very little ("only in the rare special case"), and even that might not be true yet (but we should make it true for specific tools if so). I focus on this special case because it's tenable. What you say about non-atomicity of those syscalls is not a security issue when the directory and all parent directories are trusted. It can still be a reliability and a safety issue e.g. if two sysadmins try to edit a file, but I thought that was beyond scope of our discussion.
I doubt this belongs to "SECURITY section of man pages" because this is by no means limited to just text editors. Most tools are unsafe to use on files in untrusted directories, with very few exceptions - for example, "cp" and "mv" are generally unsafe, but "ln" is generally safe.But also those tools seem not state, how they really behave regarding security in man-pages, declaring what security expectations they fulfil and which the will not fulfil (I searched for security/concurrent/user/owner/privileg but no relevant hits in the man page). How should a normal user know the difference?
I am not saying things are good as they are; I think they are not. Like I say, people neither know nor want to know this, and it means they continue to do things insecurely. I don't currently have a solution.
It is tricky to access files in an untrusted directory safely. Programs that knowingly do it end up using O_EXCL or O_NOFOLLOW|O_NOCTTY and such, and doing various *stat() calls, and even that is sometimes not enough. It'd be naive to expect the same from every other program accepting an arbitrary pathname.From my point of view, this mandates something like a "libSecureOpen" (trying to get that into libc as first step might be in vain), which has a solid implementation also considering different UNIX-system peculiarities and should be used by open source software doing that kind of risky operations.
IIRC, something like this was proposed in 1990s, albeit not for that extensive a use. You say "risky operations", but under the threat model you imply (root using almost any tool on pathnames with components writable by a user) almost all filesystem accesses are risky. To partially achieve what you seem to want to achieve, almost all uses of open(2) and fopen(3), etc. would need to be replaced with "secure" alternatives, and that would be bad in many ways, including breaking of customary behavior of traditional Unix command-line programs, which existing scripts rely on. We could proceed with introduction of isatty(3) and env var checks, but this would get messy. I say "partially" because there's no way for a program to know that the file it's looking at is still the file the user had looked at when they decided to run the program against that pathname. Not only the file itself could have been replaced, but an upper directory could have been. I included some steps to deal with this in the example referenced in my previous message, and one of the steps is a double-check by the user themselves after having created a hard link in a trusted directory. I suppose some alternate OS could introduce a paradigm where a user's view of the filesystem would be frozen when they stat() a file or list a directory and unfrozen after they've accessed a file in there. This is another can of worms. I guess it's more realistically (or less unrealistically) done for one thread in a program (with each thread having its own filesystem view freeze) rather than for a user's shell running multiple programs one after another.
Other software should explicitely declare: "is not safe for operating on file of different users/NFS in untrusted environments".
This is true for 99%+ of Unix software. Exceptions are few (like some uses of "ln", and even then there's the issue of parent directories). Alexander
Current thread:
- nvi crash recovery (was Re: [oss-security] Re: Security risk of server side text editing in general and vim.tiny specifically), (continued)
- nvi crash recovery (was Re: [oss-security] Re: Security risk of server side text editing in general and vim.tiny specifically) Hanno Böck (Nov 03)
- Re: nvi crash recovery Jakub Wilk (Nov 03)
- Re: nvi crash recovery Jakub Wilk (Nov 04)
- Re: nvi crash recovery (was Re: [oss-security] Re: Security risk of server side text editing in general and vim.tiny specifically) Daniel Micay (Nov 03)
- nvi crash recovery (was Re: [oss-security] Re: Security risk of server side text editing in general and vim.tiny specifically) Hanno Böck (Nov 03)
- Re: Re: Security risk of server side text editing in general and vim.tiny specifically Christos Zoulas (Nov 03)
- AW: Re: Security risk of server side text editing in general and vim.tiny specifically Fiedler Roman (Nov 06)
- Re: Security risk of server side text editing in general and vim.tiny specifically Solar Designer (Nov 13)
- AW: Security risk of server side text editing in general and vim.tiny specifically Fiedler Roman (Nov 13)
- Re: Security risk of server side text editing in general and vim.tiny specifically Fiedler Roman (Nov 03)
- Re: Security risk of server side text editing in general and vim.tiny specifically Fiedler Roman (Nov 03)
- Re: Security risk of server side text editing in general and vim.tiny specifically Solar Designer (Nov 03)
- Re: Security risk of server side text editing in general and vim.tiny specifically Solar Designer (Nov 03)
- Re: Security risk of server side text editing in general and vim.tiny specifically Leonid Isaev (Nov 05)
- Re: Security risk of server side text editing in general and vim.tiny specifically Solar Designer (Nov 03)