This thesis presents the design, implementation, and evaluation of a novel capability operating system: CheriOS. The guiding motivation behind CheriOS is to provide strong security guarantees to programmers, even allowing them to continue to program in fast, but typically unsafe, languages such as C. Furthermore, it does this in the presence of an extremely strong adversarial model: in CheriOS, every compartment -- and even the operating system itself -- is considered actively malicious. Building on top of the architecturally enforced capabilities offered by the CHERI microprocessor, I show that only a few more capability types and enforcement checks are required to provide a strong compartmentalisation model that can facilitate mutual distrust. I implement these new primitives in software, in a new abstraction layer I dub the nanokernel. Among the new OS primitives I introduce are one for integrity and confidentiality called a Reservation (which allows allocating private memory without trusting the allocator), as well as another that can provide attestation about the state of the system, a Foundation (which provides a key to sign and protect capabilities based on a signature of the starting state of a program). I show that, using these new facilities, it is possible to design an operating system without having to trust the implementation is correct.

CheriOS is fundamentally fail-safe; there are no assumptions about the behaviour of the system, apart from the CHERI processor and the nanokernel, to be broken. Using CHERI and the new nanokernel primitives, programmers can expect full isolation at scopes ranging from a whole program to a single function, and not just with respect to other programs but the system itself. Programs compiled for and run on CheriOS offer full memory safety, both spatial and temporal, enforced control flow integrity between compartments and protection against common vulnerabilities such as buffer overflows, code injection and Return-Oriented-Programming attacks. I achieve this by designing a new CHERI-based ABI (Application Binary Interface) which includes a novel stack structure that offers temporal safety. I evaluate how practical the new designs are by prototyping them and offering a detailed performance evaluation. I also contrast with existing offerings from both industry and academia.

CHERI capabilities can be used to restrict access to system resources, such as memory, with the required dynamic checks being performed by hardware in parallel with normal operation. Using the accelerating features of CHERI, I show that many of the security guarantees that CheriOS offers can come at little to no cost. I present a novel and secure IO/IPC layer that allows secure marshalling of multiple data streams through mutually distrusting compartments, with fine-grained authenticated access control for end-points, and without either copying or encryption. For example, CheriOS can restrict its TCP stack from having access to packet contents, or restrict an open socket to ensure data sent on it to arrives at an endpoint signed as a TLS implementation. Even with added security requirements, CheriOS can perform well on real workloads. I showcase this by running a state-of-the-art webserver, NGINX, atop both CheriOS and FreeBSD and show improvements in performance ranging from 3x to 6x when running on a small-scale low-power FPGA implementation of CHERI-MIPS.