MadMode

Dan Connolly's tinkering lab notebook

My Capability Security 2017 Wish-List

Computers are getting faster, smaller, more connected, and more capable, but when it comes to security, everything is broken. Along with correct-by-construction software (e.g. Certified Programming with Dependent Types), the best weapon I see is object capability discipline.

Before I get into my wish list of projects and issues, I'd like to point out dckc/awesome-ocap, my list of capability security technology that is ready to use today, including everything from seL4, an open source operating-system kernel with an end-to-end proof of implementation correctness and security enforcement to Sandstorm, a self-hostable SAAS platform.

Secure module loading for node.js

The good parts of JavaScript line up well with object capability discipline, but support in node.js for some of the best parts is lacking and hence there's no guarantee that calling and enhanced sqrt() from some npm module will not send an HTTP request to launch missiles.

Mark Miller demonstrated the feasibility of secure loading as far back as 2011 with makeSimpleAMDLoader.js. I'm trying to fully understand node's incomplete support for Object.defineProperty in ses/issues/6.

Meanwhile, I'm having fun with Capper; see finquick.

Sandstorm dev tools on Ubuntu 16.04

Ubuntu's 16.04 kernels handle pid namespaces in a way that interferes with the sandstorm dev tools. Tracking issue: 2526.

I figured out how to build sandstorm apps with nix, but without the dev tools, the edit/test/debug cycle time is too long.

Capability security for mainstream linux with CloudABI, capsicum

While considering alternative kernels, I ran into a linking issue when trying to build a linux kernel that supports capsicum and CloudABI. (There was a PPA for capsicum for a while...)

CloudABI or capsicum at work would be so great. But it's a long way off... we're struggling to migrate to SuSE 12 so we can try out Docker.

For example, a research workflow app I maintain needs to be able to send mail, but - only from one address - only using templated bodies - only to users who have in some way asked for it

Design sketch: at at investigator request time, user grants "capability to send app-template mail to addresses X, Y, Z".

As a demonstration, I'm porting ZeroVault to CloudABI using a FreeBSD vagrant box VM. It's pretty fun since Ed fixes my issues within a few hours of when I report them.

bus1, Capability-based IPC for Linux

I'm heartened by momentum around bus1.

On top of the lack of composability in the chmod/chgroup, there's a mounting kludge tower of stuff like SELinux and (to a lesser extent?) AppArmour. I was doing a storage audit and learning how lsblk gets the serial number of my drives. I had heard of udev and systemd, but I had no idea it uses netlink ("a more flexible alternative to ioctl") to communicate with the kernel.

Object capability discipline for Docker

Is this even possible? I can't get my head around the Docker security story.

Uniform, composable FFI and stdlib for pony

Pony aims to be a high-performance capability-secure language. I would love to see it make some inroads on golang: while go addresses (many of) the memory-safety issues of C/C++, its standard library is full of ambient authority and its type system dooms developers to lots of boilerplate maintenance.

I'm struggling (mostly for time) to convince the pony community that a reasonably simple policy can eradicate ambient authority from the standard library.

In discussion of my network API PR, I learned that the pony designers don't (yet) see interposition as a key component of robust composition.

Safe systems programming on seL4 and genode

The genode May 2016 release included initial support for rust. I haven't managed to try it out. Support for pony on genode has only gotten as far as a Dec 2015 twitter exchange as far as I know.

Robigalia aims to be a persistent capability OS built on seL4 and rust.

Rust on seL4 is pretty bleeding-edge: SEL4PROJ/rust-camkes-samples/issues/1 documents my trials and tribulations. https://github.com/seL4/refos looks interesting.

Object capability discipline support in rust

We supporters of the 2012 proposals to isolate ambient authority in the rust stdlib didn't make our case well enough for the 1.0 cut-off, but there is renewed interst in refactoring std for ultimate portability. One result of this could be a std alternative with no ambient authority.