It’s also hard to overstate how obscure Navajo was. Young and Morgan developed a writing system for it in 1937, and the first dictionary was developed as part of the war effort (and thus not publicly available) in the 1940s. The Navy estimated that ~30 non-Navajos spoke the language at the time.
It’s thus pretty unlikely that “this is Navajo” would have been an Axis analyst’s first guess—-and even if it were, there were minimal resources available. The language is pretty far from Japanese or German, which also doesn’t help.
On top of all that, the messages being sent weren’t literal Navajo—-there was sort of a substitution code that used common Navajo words to stand in for military items. The Japanese actually captured a [Navajo solider](https://en.m.wikipedia.org/wiki/Joe_Kieyoomia) but he could not undertatand the transmitted messages.
They could also have “fun” and make it very hard to crack. For example they could call a submarine a “metal fish” one day and a “sea snake” the next day and leave everybody scratching their heads. As a native speaker they could use subtle context to convey the message and that turns out to be extremely difficult for an outsider to grasp.
Since "language as cryptography" breaks Kerchoff's Principle, doesn't that just make it bad cryptography? Shouldn't that make it all the easier to use cryptanalysis to translate it?
Unrelated point: I thought Steganography was the art of hiding the _fact_ of encryption. If a Japanese analyst hears someone talking in Navajo, it's no secret that some unknown message is being conveyed.