In the decade between 1964 and 1974, several protection architectures were implemented as research and development projects, usually starting with a computer that provided only a privileged mode, adding minor hardware features and interpreting with software the desired protection architecture. Among these were M.l.T.’s CTSS which, in 1961, implemented user authentication with all-or-nothing sharing and, in 1965, added shared files with permission lists . In 1967, the ADEPT system of the System Development Corporation implemented in software on an IBM System/360 a model of the U.S. military security system, complete with clearance levels, compartments, need-to-know, and centralized authority control . At about the same time, the IBM Cambridge Scientific Center released an operating system named CP/67, later marketed under the name VM/370, that used descriptor-based hardware to implement virtual System/360 computers using a single System/360 Model 67 . In 1969, the University of California (at Berkeley) CAL system implemented a software-interpreted capability system on a Control Data 6400 computer . Also in 1969, the Multics system, a joint project of M.I.T. and Honeywell, implemented in software and hardware a complete descriptor-based access control list system with hierarchical control of authorization on a Honeywell 645 computer system , . Based on the plans for Multics, the Hitachi Central Research Laboratory implemented a simplified descriptor-based system with hardware-implemented ordered domains (rings of protection) on the HITAC 5020E computer in 1968 . In 1970, the Berkeley Computer Corporation also implemented rings of protection in the BCC 500 computer . In 1973, a hardware version of the idea of rings of protection together with automatic argument address validation was implemented for Multics in the Honeywell 6180 . At about the same time, the Plessey Corporation announced a telephone switching computer system, the Plessey 250 , based on a capability architecture.
Current experimentation with new protection architectures is represented by the CAP system being built at Cambridge University  and the HYDRA system being built at Carnegie-Mellon University  . Recent research reports by Schroeder , Rotenberg , Spier et al. , and Redell  propose new architectures that appear practical to implement.
A research problem attracting much attention today is how to certify the correctness of the design and implementation of hardware and software protection mechanisms. There are actually several sub-problems in this area.
a) One must have a precise model of the protection goals of a system against which to measure the design and implementation. When the goal is complete isolation of independent users, the model is straightforward and the mechanisms of the virtual machine are relatively easy to match with it. When controlled sharing of information is desired, however, the model is much less clear and the attempt to clarify it generates many unsuspected questions of policy. Even attempts to model the well-documented military security system have led to surprisingly complex formulations and have exposed formidable implementation problems ,  .
b) Given a precise model of the protection goals of a system and a working implementation of that system, the next challenge is to verify somehow that the presented implementation actually does what it claims. Since protection functions are usually a kind of negative specification, testing by sample cases provides almost no information. One proposed approach uses proofs of correctness to establish formally that a system is implemented correctly. Most work in this area consists of attempts to extend methods of proving assertions about programs to cover the constructs typically encountered in operating systems  .
c) Most current systems present the user with an intricate interface for specifying his protection needs. The result is that the user has trouble figuring out how to make the specification and verifying that he requested the right thing. User interfaces that more closely match the mental models people have of information protection are needed.
d) In most operating systems, an unreasonably large quantity of “system” software runs without protection constraints. The reasons are many: fancied higher efficiency, historical accident, misunderstood design, and inadequate hardware support. The usual result is that the essential mechanisms that implement protection are thoroughly tangled with a much larger body of mechanisms, making certification impossibly complex. In any case, a minimum set of protected supervisor functions—a protected kernel—has not yet been established for a full-scale modern operating system. Groups at M.l.T.  and at Mitre ,  are working in this area.
Most modern operating systems are vulnerable in their reaction to hardware failures. Failures that cause the system to misbehave are usually easy to detect and, with experience, candidates for automatic recovery. Far more serious are failures that result in an undetected disabling of the protection mechanisms. Since routine use of the system may not include attempts to access things that should not be accessible, failures in access-checking circuitry may go unnoticed indefinitely. There is a challenging and probably solvable research problem involved in guaranteeing that protection mechanisms are invulnerable in the face of all single hardware failures. Molho  explored this topic in the IBM System 360/Model 50 computer and made several suggestions for its improvement. Fabry  has described an experimental “complete isolation” system in which all operating system decisions that could affect protection are duplicated by independent hardware and software.
Another area of research concerns constraining the use to which information may be put after its release to an executing program. In Section 1, we described such constraints as a fifth level of desired function. For example, one might wish to “tag” a file with a notation that any program reading that file is to be restricted forever after from printing output on remote terminals located outside the headquarters building.
For this restriction to be complete, it should propagate with all results created by the program and into other files it writes. Information use restrictions such as these are common in legal agreements (as in the agreement between a taxpayer and a tax return preparing service) and the problem is to identify corresponding mechanisms for computer systems that could help enforce (or detect violations of) such agreements. Rotenberg explored this topic in depth  and proposed a “privacy restriction processor” to aid enforcement.
A potentially powerful technique for protecting information is to encipher it using a key known only to authorized accessors of the information. (Thus encipherment is basically a ticket-oriented system.) One research problem is how to communicate the keys to authorized users. If this communication is done inside the computer system, schemes for protecting the keys must be devised. Strategies for securing multinode computer communication networks using encipherment are a topic of current research; Branstad has summarized the state of the art  . Another research problem is development of encipherment techniques (sometimes called privacy transformations) for random access to data. Most well-understood enciphering techniques operate sequentially on long bit streams (as found in point-to-point communications, for example). Techniques for enciphering and deciphering small, randomly selected groups of bits such as a single word or byte of a file have been proposed, but finding simple and fast techniques that also require much effort to cryptanalyze (that is, with high work factors) is still a subject for research. A block enciphering system based on a scheme suggested by Feistel was developed at the IBM T. J. Watson Research Laboratory by Smith, Notz, and Osseck . One special difficulty in this area is that research in encipherment encounters the practice of military classification. Since World War II, only three papers with significant contributions have appeared in the open literature , , ; other papers have only updated, reexamined, or rearranged concepts published many years earlier.
Finally, spurred by the need for better credit and check cashing authentication, considerable research and development effort is going into better authentication mechanisms. Many of these strategies are based on attempts to measure some combination of personal attributes, such as the dynamics of a handwritten signature or the rhythm of keyboard typing. Others are directed toward developing machine-readable identification cards that are hard to duplicate.
Work in progress is not well represented by published literature. The reader interested in further information on some of the current research projects mentioned may find useful the proceedings of two panel sessions at the 1974 National Computer Conference , , a recent workshop , and a survey paper .