Demi Marie Obenour writes: > On 7/21/26 12:05, Alyssa Ross wrote: >> Demi Marie Obenour writes: >> >>> On 7/15/26 14:54, Alyssa Ross wrote: >>>> Demi Marie Obenour writes: >>>> >>>>> On 7/13/26 05:04, Alyssa Ross wrote: >>>>>> Demi Marie Obenour writes: >>>>>> >>>>>>> This seems to be a fairly significant hole in the VMM's sandbox. >>>>>>> Various code operates in that directory and isn't secure against >>>>>>> symlink attacks. >>>>>>> >>>>>>> Does the VMM really need write access to that directory? Could it >>>>>>> be given access to a subdirectory that is mounted "nosymfollow"? >>>>>> >>>>>> I think it's just so it can create e.g. the serial file. That could >>>>>> indeed be moved to a subdirectory. The ideal solution is of course for >>>>>> Cloud Hypervisor to be able to accept every file it would otherwise >>>>>> create as an open file descriptor, which I expect will happen >>>>>> eventually. >>>>> >>>>> Should it be moved to a subdirectory for now? That would make the >>>>> sandbox meaningful in practice. >>>> >>>> I suppose it could be. Might make more sense to dedicate efforts >>>> towards fixing Cloud Hypervisor to accept file descriptors for the ~2 >>>> paths it currently needs to create, though. (That's obviously more >>>> work, but also more useful.) If we were to work around this in >>>> Spectrum, we could keep symlinks outside the directory pointing into it, >>>> to keep paths nice and avoid churn once we can put them back in the >>>> parent directory. >>> >>> Do you happen to know which paths these are? >> >> I think it might just be "serial"? Not sure. > > How difficult would it be to modify Cloud Hypervisor so this is passed > as a file descriptor? I don't expect it would be very hard. It's recently been done for e.g. VFIO FDs, but I also don't think I have any special knowledge here that you don't?