On Mon, Jun 10, 2024 at 08:00:57AM GMT, Serge E. Hallyn wrote: > > Now, one thing that does occur to me here is that there is a > very mild form of sendmail-capabilities vulnerability that > could happen here. Unpriv user joe can drop CAP_SYS_ADMIN > from cap_userns, then run a setuid-root program which starts > a container which expects CAP_SYS_ADMIN. This could be a > shared container, and so joe could be breaking expected > behavior there. > > I *think* we want to say we don't care about this case, but > if we did, I suppose we could say that the normal cap raise > rules on setuid should apply to cap_userns? > Right, good catch. If we do want to fix it, we could just check for setuid no? Or do we want to follow the normal root inheritance rules too? Essentially something like this: pU' = is_suid(root) ? X : pU
