bug#21702: shell-quote-argument semantics and safety

[Eli Zaretskii <eliz@gnu.org>]

> From: taylanbayirli@gmail.com (Taylan Ulrich Bayırlı/Kammer)
> Cc: 21702@debbugs.gnu.org
> Date: Sun, 18 Oct 2015 21:12:34 +0200
> 
> I'd like to point out that (in the most extreme cases) people have
> actually been writing web servers and other such programs in Elisp for
> which one would normally use a general-purpose language.
> 
> That is, "APIs that could be maliciously abused" is not the right way to
> look at it.  It's not about the Elisp programmer abusing the API, it's
> about a malicious data source exploiting a (potential) flaw in an Elisp
> function, which Elisp programmers have relied on and thus made their
> programs vulnerable to code injection.
> 
> 
> That's why I was being so careful with regard to the safety guarantees
> of the "shell-quasiquote" package I contributed.  I would like people to
> be able to use that as part of a general-purpose Elisp language, and so
> being safe against code injection is an absolute must.  They might after
> all use it as part of a network-facing service.
> 
> 
> Actually that might also apply when using e.g. TRAMP, which also
> communicates with remote hosts and is a normal part of Emacs.  I've been
> told it receives file names from remote hosts and passes them through
> shell-quote-argument before giving them to a shell.  So maybe my
> concerns apply there as well.
> 
> 
> Given that, "I think 1) is now covered" is not very relieving to hear.

Item 1 was this:

> >> The function should clearly document
> >> 
> >>     1) for which shells will the quoting work absolutely, i.e. lead to
> >>     the given string to appear *verbatim* in an element of the ARGV of
> >>     the called command,

There's nothing about safety here, only about correctness.  That is
the aspect that I think is now covered, as the doc string now says for
which shells one can have correct results.

> It amounts to "I think this is safe against code injection" which is
> rather alarming to hear.  Either it's very confidently known to be safe
> and so one may use it for network-facing code, or it's not confidently
> known to be safe and so one shouldn't use it for network-facing code.
> This should be documented clearly especially so that users who aren't
> very aware of injection attacks won't nonchalantly use the function for
> their network-facing code (when the function isn't known to be safe for
> this), but also so that users who are aware of such issues know they can
> use the function and don't instead invent their own thing (when it is
> known to be safe).
> 
> Does that make sense?

Maybe it does, but only if we start documenting these aspects
project-wide.  It makes little sense to me to do that for a single
API, and not an important one at that.  But that's me.