Before there is further evolution of my statements on the subject, I should probably clarify what's going on.
In "The Risks of Nanotechnology" published in "Nanotechnology, Research and Perspectives" (B.C. Crandall and James Lewis, editors, MIT press 1992), I argued that the development of (a) artificial self-replicating systems (b) able to function in a natural environment and which are (c) designed to evolve; appears both very difficult and extraordinarily dangerous.
It would seem prudent, when the ability to develop such systems is imminent, to be prepared to establish an appropriate regulatory framework to prevent the rather obvious risks that are in principle possible. This would likely include mechanisms to prevent the development and uncontrolled release of such systems.
In general, when the concept of self-replicating systems is raised, people invariably think of biological self-replicating systems. This is only natural, as these provide working examples of such systems. However, biological systems have many other (often amazing) properties: they are, in particular, highly flexible and able to adapt to a broad range of circumstances. Typical manufactured systems are highly inflexible and are unable to adapt to any significant changes in their environment (unless the ability to adapt to certain types of environmental changes has been painstakingly designed in from the start).
(Another discussion of the likely differences between biological
systems and artificial systems is "Self replicating systems and low
cost manufacturing," to appear in: The ultimate limits of fabrication
and measurement, M.E. Welland, J.K. Gimzewski, eds.; Kluwer,
Dordrecht, 1994. Drexler has discussed the differences between
"M-style" (machine-style) and "O-style" (organic-style) system in
"Biological and Nanomechanical Systems: Contrasts in Evolutionary
Capacity" in "Artificial Life", edited by Christopher Langton,
Addison-Wesley, 1988).
One type of inflexibility is found with respect to the source of energy. Take, for example, a car and a horse. Both provide transportation. The horse can use any of a broad range of possible fuel sources for energy, including: straw, carrots, hay, grass, potatoes, sugar lumps, etc. The car, however, must be provided with gasoline -- a very specialized requirement that is common in the design of cars. There are several reasons for this: designing a car that could efficiently use many different types of fuel would be more complex. The manufacture of such a car would be more difficult, and the cost would no doubt be higher. Designing the car to use a single type of refined fuel, a fuel that is widely and economically available, is a very efficient and economical design. The image of a car running feral in the woods and sucking sap from trees for its energy seems highly implausible. Cars simply aren't designed to be very flexible with regard to energy source, and there are many sensible reasons for this.
While an artificial self-replicating system will indeed be able to self-replicate (like a horse) it is unlikely that it will share other properties of biological systems (e.g., their marvelous flexibility and adaptibility to new and unpredictably changing environments). Imagining that artificial self-replicating systems designed for the economical manufacture of products will necessarily have all the other properties of biological systems leads to the "Star Trek" scenario, in which two "nanites" (with the delicate implication of sexual reproduction) escape into the computer, find that parts in the computer are (for entirely inexplicable reasons) suitable "food" for nanites, evolve into an intelligent species and, by the episodes end, successfully negotiate for their own planet.
This scenario does not appear plausible....
The deliberate design of an artificial self-replicating system able to function in a natural environment and capable of evolution does not, however, appear infeasible. As a consequence, deliberate abuse of the technology would seem to be of greater concern than accidents.
In any event, using the principles of evolution to aid in the design of systems; provided we aren't engaged in activites that pose extraordinary risks (such as indiscriminately releasing robust, evolveable self-replicating systems into the environment); would seem to be entirely reasonable.