The earliest known use of cryptography is found in non-standard hieroglyphs carved into monuments from Egypt's Old Kingdom (ca 4500 years ago). These are not thought to be serious attempts at secret communications, however, but rather to have been attempts at mystery, intrigue, or even amusement for literate onlookers. These are examples of still another use of cryptography, or of something that looks (impressively if misleadingly) like it. Later, Hebrew scholars made use of simple Substitution ciphers (such as the Atbash cipher) beginning perhaps around 500 to 600 BCE. Cryptography has a long tradition in religious writing likely to offend the dominant culture or political authorities. Perhaps the most famous is the 'Number of the Beast' from the book of Revelations in the Christian New Testament. '666' is almost certainly a cryptographic (i.e., encrypted) way of concealing a dangerous reference; many scholars believe it's a concealed reference to the Roman Empire, or the Emperor Nero, (and so to Roman policies of persecution of Christians) that would have been understood by the initiated (who 'had the codebook'), and yet be safe (or at least somewhat deniable and so less dangerous) if it came to the attention of the authorities. At least for orthodox Christian writing, the need for such concealment ended with Constantine's conversion and the adoption of Christianity as the official religion of the Empire.
The Greeks of Classical times are said to have known of ciphers (e.g., the scytale transposition cypher claimed to have been used by the Spartan military). Herodutus tells us of secret messages physically concealed beneath wax on wooden tablets or as a tattoo on a slave's head concealed by regrown hair (these are not properly examples of cryptography per se; see secret writing). The Romans certainly did (e.g., the Caesar cipher and its variations). There is ancient mention of a book about Roman military cryptography (especially Julius Caesar's); it has been, unfortunately, lost.
In India, cryptography was apparently well known. It is recommended in the Kama Sutra as a technique by which lovers can communicate without being discovered. This may imply that cryptanalytic techniques were less than well developed in India ca 500 CE.
Cryptography became (secretly) important still later as a consequence of political competition and religious analysis. For instance, in Europe during and after the Renaissance, citizens of the various Italian states, including the Papacy, were responsible for substantial improvements in cryptographic practice (e.g., polyalphabetic ciphers invented by Leon Alberti ca 1465). And in the Arab world, religiously motivated textual analysis of the Koran led to the invention of the frequency analysis technique for breaking monoalphabetic substitution cyphers sometime around 1000 CE.
Cryptography, cryptanalysis, and secret agent betrayal featured in the Babington plot during the reign of Queen Elizabeth I which led to the execution of Mary, Queen of Scots. And an encrypted message from the time of the Man in the Iron Mask (decrypted around 1900 by Étienne Bazeries) has shed some, regrettably non-definitive, light on the identity of that legendary, and unfortunate, prisoner. Cryptography, and its misuse, was involved in the plotting which led to the execution of Mata Hari and even more reprehensibly, if possible, in the travesty which led to Dreyfus' conviction and imprisonment, both in the early 20th century. Fortunately, cryptographers were also involved in setting Dreyfus free; Mata Hari, in contrast, was shot.
Mathematical cryptography leapt ahead (also secretly) after World War I. Marian Rejewski, in Poland, attacked and 'broke' the early German Army Enigma system (an electromechanical rotor cypher machine) using theoretical mathematics in 1932. The break continued up to '39, when changes in the way the German Army's Enigma machines were used required more resources than the Poles could deploy. His work was extended by Alan Turing, Gordon Welchman, and others at Bletchley Park beginning in 1939, leading to sustained breaks into several other of the Enigma variants and the assorted networks for which they were used. US Navy cryptographers (with cooperation from British and Dutch cryptographers after 1940) broke into several Japanese Navy crypto systems. The break into one of them famously led to the US victory in the Battle of Midway. A US Army group, the SIS, managed to break the highest security Japanese diplomatic cipher system (an electromechanical 'stepping switch' machine called Purple by the Americans) even before WWII began. The Americans referred to the intelligence resulting from cryptanalysis, perhaps especially that from the Purple machine, as 'Magic'. The British eventually settled on 'Ultra' for intelligence resulting from cryptanalysis, particularly that from message traffic enciphered by the various Enigmas. An earlier British term for Ultra had been 'Boniface'.
World War II Cryptography
By World War II mechanical and electromechanical cryptographic cipher machines were in wide use, but they were impractical manual systems. Great advances were made in both practical and mathematical cryptography in this period, all in secrecy. Information about this period has begun to be declassified in recent years as the official 50-year (British) secrecy period has come to an end, as the relevant US archives have slowly opened, and as assorted memoirs and articles have been published.
The Germans made heavy use (in several variants) of an electromechanical rotor based cypher system known as Enigma. The German military also deployed several mechanical attempts at a one-time pad. Bletchley Park called them the Fish cyphers, and Max Newman and colleagues designed and deployed the world's first programmable digital electronic computer, the Colossus, to help with their cryptanalysis. The German Foreign Office began to use the one-time pad in 1919; some of this traffic was read in WWII partly as the result of recovery of some key material in South America that was insufficiently carefully discarded by a German courier.
The Japanese Foreign Office used a locally developed electrical stepping switch based system (called Purple by the US), and also used several similar machines for attaches in some Japanese embassies. One of these was called the 'M-machine' by the US, another was referred to as 'Red'. All were broken, to one degree or another by the Allies.
Other cipher machines used in WWII included the British Typex and the American SIGABA; both were electromechanical rotor designs similar in spirit to the Enigma.
The era of modern cryptography really begins with Claude Shannon, arguably the father of mathematical cryptography. In 1949 he published the paper Communication Theory of Secrecy Systems in the Bell System Technical Journal, and a little later the book Mathematical Theory of Communication with Warren Weaver. These, in addition to his other works on information and communication theory established a solid theoretical basis for cryptography and for cryptanalysis. And with that, cryptography more or less disappeared into secret government communications organizations such as the NSA. Very little work was again made public until the mid '70s, when everything changed.
1969 saw two major public (i.e., non-secret) advances. First was the DES (Data Encryption Standard) submitted by IBM, at the invitation of the National Bureau of Standards (now NIST), in an effort to develop secure electronic communication facilities for businesses such as banks and other large financial organizations. After 'advice' and modification by the NSA, it was adopted and published as a FIPS Publication (Federal Information Processing Standard) in 1977 (currently at FIPS 46-3). It has been made effectively obsolete by the adoption in 2001 of the Advanced Encryption Standard, also a NIST competition, as FIPS 197. DES was the first publicly accessible cypher algorithm to be 'blessed' by a national crypto agency such as NSA. The release of its design details by NBS stimulated an explosion of public and academic interest in cryptography. DES, and more secure variants of it (such as 3DES or TDES; see FIPS 46-3), are still used today, although DES was officially supplanted by AES (Advanced Encryption Standard) in 2001 when NIST announced the selection of Rijndael, by two Belgian cryptographers. DES remains in wide use nonetheless, having been incorporated into many national and organizational standards. However, its 56-bit key-size has been shown to be insufficient to guard against brute-force attacks (one such attack, undertaken by cyber civil-rights group The Electronic Frontier Foundation, succeeded in 56 hours -- the story is in Cracking DES, published by O'Reilly and Associates). As a result, use of straight DES encryption is now without doubt insecure for use in new crypto system designs, and messages protected by older crypto systems using DES should also be regarded as insecure. The DES key size (56-bits) was thought to be too small by some even in 1976, perhaps most publicly Whitfield Diffie. There was suspicion that government organizations even then had sufficient computing power to break DES messages and that there may be a back door due to the lack of randomness in the 'S' boxes.
Second was the publication of the paper New Directions in Cryptography by Whitfield Diffie and Martin Hellman. This paper introduced a radically new method of distributing cryptographic keys, which went far toward solving one of the fundamental problems of cryptography, key distribution. It has become known as Diffie-Hellman key exchange. The article also stimulated the almost immediate public development of a new class of enciphering algorithms, the asymmetric key algorithms.
Prior to that time, all useful modern encryption algorithms had been symmetric key algorithms, in which the same cryptographic key is used with the underlying algorithm by both the sender and the recipient who must both keep it secret. All of the electromechanical machines used in WWII were of this logical class, as were the Caesar and Atbash cyphers and essentially all cypher and code systems throughout history. The 'key' for a code is, of course, the codebook, which must likewise be distributed and kept secret.
Of necessity, the key in every such system had to be exchanged between the communicating parties in some secure way prior to any use of the system (the term usually used is 'via a secure channel') such as a trustworthy courier with a briefcase handcuffed to a wrist, or face-to-face contact, or a loyal carrier pigeon. This requirement rapidly becomes unmanageable when the number of participants increases beyond some (very!) small number, or when (really) secure channels aren't available for key exchange, or when, as is sensible crypto practice keys are changed frequently. In particular, a separate key is required for each communicating pair if no third party is to be able to decrypt their messages. A system of this kind is also known as a private key, secret key, or conventional key cryptosystem. D-H key exchange (and succeeding improvements) made operation of these systems much easier, and more secure, than had ever been possible before.
In contrast, with asymmetric key encryption, there is a pair of mathematically related keys for the algorithm, one of which is used for encryption and the other for decryption. Some, but not all, of these algorithms have the additional property that one of the keys may be made public since the other cannot be (by any currently known method) deduced from the 'public' key. The other key in these systems is kept secret and is usually called, somewhat confusingly, the 'private' key. An algorithm of this kind is known as a public key / private key algorithm, although the term asymmetric key cryptography is preferred by those who wish to avoid the ambiguity of using that term for all such algorithms, and to stress that there are two distinct keys with different secrecy requirements.
As a result, for those using such algorithms, only one key pair is now needed per recipient (regardless of the number of senders) as possession of a recipient's public key (by anyone whomsoever) does not compromise the 'security' of messages so long as the corresponding private key is not known to any attacker (effectively, this means not known to anyone except the recipient). This unanticipated, and quite surprising, property of some of these algorithms made possible, and made practical, widespread deployment of high quality crypto systems which could be used by anyone at all. Which in turn gave government crypto organizations worldwide a severe case of heartburn; for the first time ever, those outside that fraternity had access to cryptography that wasn't readily breakable by the 'snooper' side of those organizations. Considerable controversy, and conflict, began immediately. It has not yet subsided. In the US, for example, exporting strong cryptography remains illegal; cryptographic methods and techniques are classified as munitions. Until 2001 'strong' crypto was defined as anything using keys longer than 40 bits -- the definition was relaxed thereafter. (See S Levy's Crypto for a journalistic account of the policy controversy in the US).
Note, however, that it has NOT been proven impossible, for any of the good public/private asymmetric key algorithms, that a private key (regardless of length) can be deduced from a public key (or vice versa). Informed observers believe it to be currently impossible (and perhaps forever impossible) for the 'good' asymmetric algorithms; no workable 'companion key deduction' techniques have been publicly shown for any of them. Note also that some asymmetric key algorithms have been quite thoroughly broken, just as many symmetric key algorithms have. There is no special magic attached to using algorithms which require two keys.
In fact, some of the well respected, and most widely used, public key / private key algorithms can be broken by one or another cryptanalytic attack and so, like other encryption algorithms, the protocols within which they are used must be chosen and implemented carefully to block such attacks. Indeed, all can be broken if the key length used is short enough to permit practical brute force key search; this is inherently true of all encryption algorithms using keys, including both symmetric and asymmetric algorithms.
This is an example of the most fundamental problem for those who wish to keep their communications secure; they must choose a crypto system (algorithms + protocols + operation) that resists all attack from any attacker. There being no way to know who those attackers might be, nor what resources they might be able to deploy, nor what advances in cryptanalysis (or its associated mathematics) might in future occur, users may ONLY do the best they know how, and then hope. In practice, for well designed / implemented / used crypto systems, this is believed by informed observers to be enough, and possibly even enough for all(?) future attackers. Distinguishing between well designed / implemented / used crypto systems and crypto trash is another, quite difficult, problem for those who are not themselves expert cryptographers. It is even quite difficult for those who are.
Revision of modern history
In recent years public disclosure of secret documents held by the UK government has shown that asymmetric key cryptography, D-H key exchange, and the best known of the public key / private key algorithms (i.e., what is usually called the RSA algorithm), all seem to have been developed at a UK intelligence agency before the public announcement by Diffie and Hellman in '76. GCHQ has released documents claiming that they had developed public key cryptography before the publication of Diffie and Hellman's paper. Various classified papers were written at GCHQ during the 1960s and 1970s which eventually led to schemes essentially identical to RSA encryption and to Diffie-Hellman key exchange in 1973 and 1974. Some of these have now been published, and the inventors (James Ellis, Clifford Cocks, and Malcolm Williamson) have made public (some of) their work and stuff.