As core Capsicum libraries will appear in FreeBSD 9 anyway, I think it's possible to take several applications from the base system and modify them to use Capsicum sandboxes. For example, the FreeBSD syslog daemon might be an interesting application to adapt to compartmentalisation model. Currently I see the following milestones for this project:
1. Determine the current state of Capsicum integration in FreeBSD – obviously nessesary to know. This includes compiling the respective Perforce branch to see what's going on there. Also I plan to get in touch with people involved in implementing initial Capsicum prototype for FreeBSD. This may be done during the Community Bonding Period, as this doesn't involve writing any code and hence doesn't break GSoC's rules. DONE
2. Build a list of applications that should be adapted to use enhanced security model. I've received some suggestions that applications interacting with network should be changed first. This includes various system services (syslogd, sshd, DNS resolver, inetd and it's slave servers). Requires discussion with $mentor though. DONE -- there was a discussion on hackers@
3. For each of such applications, discuss with hackers@/$mentor how to implement required sandboxing better. This will also highlight what features should be added to Capsicum framework itself ("services" for sandboxes?) Currently in progress
4. Based on the applications list described in the previous step, decide which of them should be implemented during this GSoC, consider every application as a milestone, implement. So people will be able to test and discuss the changes.If some application requires libcapsicum changes, do it first, announce update, discuss it, then implement application changes. Currently in progress
As security issues in FreeBSD are usually not known prior to their announcement on freebsd-security (usually with patches for correcting situation), I think that functional testing should involve playing with procstat(1) to determine opened file descriptors, etc (as in original Capsicum paper published by Google), maybe other instruments for other sets of resources.
Also it's possible to insert "malicious" code into the parts of application that are believed to be potentiallly vulnerable (compression/decompression routines, parsing code) and show that this code isn't able to do harm when operating in sandboxed environment.
Regression testing would involve measuring performance overhead introduced by capability checks.