[sweis]

It's not just two messages colliding on the exact millisecond. They can collide on subsequent blocks. So if Alice sends a 10 block message that starts at T_0 and Bob sends a 5 block message that starts at T_5, then the server learns information about half of Alice's plaintext and all of Bob's. Each client should use an independently-chosen and unpredictable IV.

You are incorrect on your statement that the corrupted server needs both the plain and encrypted versions of a message to send bogus messages. Without authentication, the server can flip arbitrary bits of a CTR message. This opens up several types of attacks. You should apply a MAC to the ciphertext.

I think that the encryption needs to happen entirely in the client and you can't rely on code downloaded from an untrusted server.

[nickbw]

Thank you for the insightful comments, and for taking the time to read the code! I really appreciate it.

I've added a pseudo-random component to the nonce, and a MAC to the messages.

I certainly agree that you can't rely on code from untrusted servers, but I think doing the encryption in the client with code that is publicly viewable, even if it can be compromised at any time by a malicious server, is the best we can do for web apps unless/until major browser vendors incorporate client-to-client encryption. Right now there is no reliable protection against a compromised server, but I would like to at least see web apps strive to be more accountable.