Re: org-crypt leaking data when encryption password is not entered twice (was: Please document the caching and its user options)

[Ihor Radchenko <yantar92@posteo.net>]

Daniel Clemente <n142857@gmail.com> writes:

>> One simple idea is to disable backups if encryption fails.
>> Or use `write-contents-functions' instead of `before-save-hook' - that
>> way, Emacs will not ignore errors thrown by org-crypt and will not
>> actually save anything if encryption fails.
>>
>
> Disabling backups makes sense too, if we decide that unencrypted
> private data shouldn't end up in backups.
> I don't have an absolute opinion. Some people may prefer having
> backups of all data (including private unencrypted data).

Actually, thinking about it more, I realize that backups may never
contain unencrypted data as long as we never write this unencrypted data
when saving normally. That's because backup is always taken from disk
and never from the buffer contents.

So, the real problem to solve is how to _reliably_ prevent the
unencrypted data to be saved onto the disk.

> If it's possible to detect whether encryption failed in this buffer,
> there could be a warning saying „Last encryption failed. Really
> save?“.

Yes. In fact, `org-entrypt-entries' throws an error when encryption
fails. However, this error is displayed as a simple message, which is
immediately hidden by "Wrote ..." message emitted a bit later.

This is because `basic-save-buffer' has

;; Don't let errors prevent saving the buffer.
(with-demoted-errors "Before-save hook error: %S"
  (run-hooks 'before-save-hook))

If we use `write-contents-functions' instead of `before-save-hook',
there should be no such problem.

> Or just a message in the style of „Encryption failed. Saving the file
> may store unencrypted data in disk, and in backups and cache if
> enabled“.
>
> Totally preventing the user from saving a file seems harsh but it also
> seems safer. Since users have different safety preferences, Emacs can
> let the user decide what the do, through a question or optional
> setting.

I agree that "saving prevention" must be a user option.

>> These things should be considered bugs. And we should fix them. Cache and
>> other libraries should not be responsible for special treatment of
>> optional org-crypt library.
>>
>
> You can't fix all bugs all the time, so you can't base security on „we
> strongly believe there are no more bugs“.

I did not suggest that.
What I am saying is that "we might have bugs, so be careful" is not
something we need to write in the documentation. The only exception is
when there is a known, long-living bug, that we cannot fix quickly and
must warn users about.

> ... If doing an extra
> verification (to avoid storing private data on disk in unencrypted
> form) is fast, it's better with the verification.

>> Cache and other libraries should not be responsible for special
>> treatment of optional org-crypt library.
>
> That's arbitrary. Both persistent cache and org-crypt are optional,
> but any of them can check whether the other is enabled and try to do
> what the user wants.
> I know they both have separate responsibilities, but if there are only
> these 2 parts, one of them must be the one caring about „unencrypted
> data leaking into disk caches“.

Sure. But I meant that we should still write this code in org-crypt
library, not inside org-persist. This is more of a technical detail and
code style.

> In addition, „leaving some encrypted sections unencrypted for a short
> amount of time, and closing and reopening the buffer during that time“
> isn't a bug, it's a possible user behaviour that we can't control. But
> org-crypt can mention that that behaviour is unsafe when using on-disk
> cache. Or detect it (if it's fast) and warn the user.

I did not mean that opening/closing buffer is a bug.
And I do not see why this behavior is unsafe, sorry.

> It would be different If we had a third component… E.g. imagine we had
> a component/overlay/text property/… in Emacs that could tell whether a
> buffer's region contains very private information or not; then all
> other components could just obey that setting (that section won't be
> backed up, it won't end up in disk cache, … It can even be displayed
> in a different face). Then org-crypt just needs to set that flag when
> encryption fails. Does something like that exist? Anyway this is a bit
> utopic or overengineered. Simpler ways of improving things are with
> documentation (e.g. „Don't do this, it's unsafe“), with messages
> („You're doing this, which may be unsafe“), or with questions („Really
> do this unsafe thing?“)

Sounds interesting, but I am afraid that this idea is too abstract. It
is not clear what Emacs is supposed to do with such regions. Maybe Eli
has something better to say.

-- 
Ihor Radchenko // yantar92,
Org mode contributor,
Learn more about Org mode at <https://orgmode.org/>.
Support Org development at <https://liberapay.com/org-mode>,
or support my work at <https://liberapay.com/yantar92>