Including pam_faildelay module in PAM distribution/possible security problem

[Bjoern Voigt]

Hello!

Thorsten Kukuk wrote:
> On Sun, Oct 22, Bjoern Voigt wrote:
>   
>> Hello!
>>
>> Since long time I miss support for PAM's pam_fail_delay() function in 
>> common Linux applications. For instance, OpenSSH and SUDO don't support 
>> fail delay without special patches.
>>     
>
> Something I should mention here: pam_fail_delay() has a huge drawback/
> security problem:
>
> Since libpam is sleeping here, the application has no chance to log
> the failed authentication attemp.
>
> Depending on the application, you can abort the application with
> Ctrl-C while it is inside of the delay function. As result, there
> will never be a syslog entry about the failed login attemp.
>
> We had some time (years?) ago the security problem that attacker 
> could use some applications for brute force attacks by aborting 
> them during the delay, because the following happend:
>
> - authentication fails
> - sleep()
> - syslog()
>
> This was changed to
> - authentication fails
> - syslog ()
> - sleep ()
>
>
> With pam_fail_delay, you will always have sleep before syslog, this is
> not changeable.
>
> And I have no good idea how to make sure that the libpam sleep will
> not be aborted.
>   
Ok, I see the problem. I also don't have a complete solution for this 
problems. But may be we find a good solution together. My comments:

   1. Not all apps can be canceled with Control-C. Especially the server
      applications should not be affected. Console programs sometimes
      avoid this situation itself (such programs like "login").
   2. If the application does not log the failure, the admin can look
      for authentication failures in PAM's syslog entries.
   3. I don't like the hardcoded "sleep" function very much. This is
      especially problematic within GUI programs. A GUI program can not
      react events if it wait's for PAM. Ideally an application could
      register a custom wait/sleep callback function. Unfortunately such
      a new callback would not help to secure unmodified programs.

Greetings,
Björn