I'm quite certain that the current TextSecure chat allows my proposed scenario with Alice, Bob and Carol to go through without issue. This is the main problem h

I'm quite certain that the current TextSecure chat allows my proposed scenario with Alice, Bob and Carol to go through without issue. This is the main problem here. So while transcript consistency is discussed in the blog post, it remains the case that Alice can send a different message to Bob and Carol without being detected. Are you sure this is correct? From the blog post, it seems like this is impossible: https://whispersystems.org/blog/images/groups-pairwise-optimize.png https://whispersystems.org/blog/images/groups-pairwise-optim... However, there’s an optimization we can make for longer messages and media. The sending client generates an ephemeral symmetric key K, encrypts the message with K (C = EK(P)), and then transmits a single copy of the ciphertext (C) along with the pairwise encryptions of the plaintext hash and the small key K: So, when a client wants to send a message to the group (like "How are you?") under this scheme, the client encrypts the message using a random encryption key K, yielding message ciphertext C. The client transmits to the server, then the server sends C to every other member of the group. Since C can only be decrypted by someone who knows K, our client must of course send K to every other member of the group. This is easily accomplished: the client encrypts K once per other member. So, if there are N members in the group, this yields N-1 small ciphertexts. Let's call them "key ciphertexts". These "key ciphertexts" are sent to the server and routed to the appropriate recipient, who decrypts it, thus receiving K. Then the recipient decrypts C using K, yielding the client's original message ("How are you?"). So even though the client is indeed generating a ciphertext per recipient, that ciphertext contains nothing but the decryption key K. It doesn't contain the client's actual message. The actual message is contained in ciphertext C, which is generated by the client and sent to the server only once, and C is relayed verbatim to every other member. That's why I say your attack seems impossible: every member has the same message ciphertext, C, so there's no opportunity for a malicious client to send different message ciphertexts to different clients.) As long as TextSecure uses this implementation of group messaging, then your attack shouldn't be possible, right?