Why not? It's like any language runtime, except way more flexible. Jonathan Rees also did some work on security kernels for it:

http://dspace.mit.edu/bitstream/handle/1721.1/36956/32890570...

Although, I think my default idea of a separation kernel plus (runtime/VM here) + (Camkes or Cap n Proto) for glue would work for LISP images, too. Throw in VLISP for as much component reliability as you get flexibility.