http://lua-users.org/wiki/SandBoxes
There's a slide titled security in Marc Balmer's talk, but I didn't get where he was headed.
Having a dynamic interpreted language runtime available as just another item in the syscall table [2] would allow arbitrary, unrestricted execution of code that would not even respect resource limitations set in the root user's shell.
[0] http://en.wikipedia.org/wiki/Mandatory_access_control
The short answer will be, it depends on how it's done, and what the scope is. You wouldn't want arbitrary code loaded in.
Having the interpreter always available in user space seems more logical.
You could do some fast and very sophisticated manipulation of network traffic streams without requiring that each packet be copied to and from user-space. You wouldn't even have to develop much packet-twiddling code. There is a surprisingly large amount of network-specific open source Lua.
Wireshark, the most popular network protocol analyzer, has dissectors written in Lua. [0]
Nmap, the network mapper and port scanner, has a Lua-based scripting engine to allow session and endpoint property discovery in great detail. [1]
The popular intrusion detection tool Snort Security Platform (SnortSP) has included a Lua interpreter since version 3.0. [2]
Yes, keeping the interpreter in user-space is safer. I have a hard time imagining sandboxing within the language runtime being able to deal with some of the nastier traffic. Perhaps one prudent thing to do with an in-kernel interpreter would be to use a project-specific sanitizer that could reliably escape tokens and keywords found in live traffic.
[0] "Lua Support in Wireshark" http://www.wireshark.org/docs/wsug_html_chunked/wsluarm.html
[1] "Nmap Scripting Engine" http://nmap.org/book/nse.html
[2] “A Pig(Snort), A Moon (Lua) and one very happy developer (Bill)” http://www.linuxsecurity.com/content/view/128135