Since 1996 there has been tremendous progress in developing the idea of certified code, including both proof-carrying code (PCC) and typed assembly language (TAL). In a certified code framework, each program (which is usually in machine-code binary form) comes equipped with a certificate that "explains," both rigorously and in a manner that is easily validated, why it possesses a formally specified safety property. A substantial amount of the research work in this area has been directed towards the problem of how to make certified code a practical technology?what one might call "proof engineering." Thus, many of the advances have been in methods for representing the certificates in the most compact and efficiently checkable way. A considerable amount of effort has also gone into the development of prototype tools that explore how to handle realistic programs written in realistic languages. In this talk, I will start with a brief overview of the current state of these and other current concepts in certified code. Then I will consider a very different but equally practical question: Just what is it that we are trying to prove, especially if we want to be relevant to computer security? Today, certified code systems do not prove the semantic equivalence between source and target programs. Nor do they prove the absence of most kinds of trojan horses, covert channels, or race conditions. While the safety properties provided by current certified code systems are, in fact, of central importance to computer security, I will argue that there are potentially great opportunities in investigating an expansion of the kinds of properties that these systems reason about.