I don't know if compression offers much protection against plaintext attacks.
This also makes me wonder how helpful AI is in such situations. AI is essential an extremely effective, lossy, compression algorithm.
> we show that it is possible to identify the phrases spoken within encrypted VoIP calls when the audio is encoded using variable bit rate codecs
https://crypto.stackexchange.com/a/2188
See also https://breachattack.com/ when the plaintext is partially attacker-controlled.
These paraphrasing instructions could be followed. But the paraphrasing could be done using some LLM. A sufficiently advanced adversary manages to invert the model somehow, and as a result can get the original plain text out of the paraphrased message, which lets them do a known-plaintext attack, get the key, and use it on other messages.
Sort of technobabble (is the idea of inverting an LLM nonsense?) but fun.