[less_less]

If you're just talking about basic encryption achieving no properties other than security against passive attack, then just nesting is probably fine. But in more complex systems, things end up being fairly subtle, so cryptographers aim for very specific security properties for the building blocks, and then try to combine these in a way that can be proved secure, at least in some attack model. This reduces the likelihood of a complex attack like the TLS triple handshake attack, where the protocol looks fine intuitively but can be broken by some weird pattern of forwarding messages between multiple parties.

So for example, nesting does not preserve IND-CCA security ("indistinguishability under chosen ciphertext attack"). Suppose you set ciphertext = outer_encrypt(outer_key, inner_encrypt(inner_key, data)). If the outer encryption system is broken, then an attacker can strip the outer layer and re-encrypt it. This will result in a different ciphertext, because if either layer is aiming for IND-CCA security, the encryption is necessarily randomized. Being able to modify ciphertexts in this way violates the IND-CCA-security goal. Then if another part of the system is designed assuming that the cipher is IND-CCA-secure, then its security is now at risk.

The attack surface is even broader if the system supports multiple combinations of ciphers, where an attacker might strip off one cipher layer and replace it with a different one.