Dozens of the largest names in crypto security, Coinbase, Block and BitGo among them, have signed a letter asking frontier AI labs for something that sounds odd until you sit with it for ten seconds: the same capabilities that attackers are already using. The argument is that safety guardrails on the best models block legitimate defensive work, vulnerability research, exploit reproduction, adversarial testing, while the people attacking these systems face no such limits. Defenders queue for permission. Attackers do not.
You can argue about whether the ask is the right fix. The diagnosis is hard to argue with. And it is not really a story about AI at all. It is a story about who can afford to check.
TL;DR
- Coinbase, Block, BitGo and dozens of other crypto firms signed a letter arguing that safety guardrails on frontier AI models block legitimate security research while attackers operate without restriction.
- The real issue is capability asymmetry: if finding out whether a system is honest requires expensive, gated, adversarial tooling, then honesty is only checkable by the well resourced.
- That is exactly how closed-RNG crypto casinos work. Detecting a rigged server-side generator requires volume, statistics and expertise most players will never have.
- THE COST OF CHECKING is the seventeenth unasked half of fairness: a fairness claim is only meaningful if verifying it is cheaper than defeating it.
- Satoshie inverts the asymmetry. A Chainlink VRF proof is verified mechanically on-chain by the coordinator before the callback fires, and any player can re-read the result on Basescan with no tooling, no permission and no arms race.
The asymmetry the letter is actually describing
Strip out the AI specifics and the complaint reduces to a shape anyone who has run a security programme will recognise. Attack is cheap and unilateral, defence is expensive and permissioned. What the letter adds is that the gap is now being widened by the tooling itself: the best instruments for probing a system are handed out selectively, and the selection process applies only to people who identify themselves and ask nicely. That is a filter on the honest, not on the hostile.
Notice what that does to verification. If the only way to establish whether something is trustworthy is to attack it competently, trustworthiness stops being a property of the system and becomes a property of the auditor’s budget. Rent the capability, get an answer. Cannot afford it, take it on faith.
The seventeenth unasked half of fairness: the cost of checking
Every instalment in this series has been about a property that sits next to provable fairness and never gets marketed. Transaction ordering. Liveness. The front end. Recourse. Settlement. The inputs nobody verified. This one is simpler than all of them and sits underneath the lot.
A fairness claim is only worth anything if verifying it is cheaper than defeating it.
That is the whole test. Not “is the system honest”, which nobody can answer in the abstract, but “what does it cost me, personally, with no special access, to find out”. If the answer is a statistics degree and six months of logged outcomes, the claim is decorative. If the answer is opening a block explorer, the claim is real.
Almost nobody in crypto gaming asks this question, because almost nobody in crypto gaming would like the answer.
How the closed casino fails the test
Take a conventional crypto casino running server-side randomness. The operator says the games are fair. Suppose they are not. What would it take you to establish that?
You would need a large sample of outcomes, which means either losing a lot of money or convincing other players to pool their histories. You would need the intended distribution to test against, which the operator has no obligation to publish honestly. You would need enough statistical competence to separate a rigged generator from ordinary variance, which is genuinely hard, because a well-implemented bias does not look like a bias at any sample size a normal player will ever reach. And at the end you would have an inference, not a proof, dismissable with one sentence about your unlucky run.
Meanwhile the operator does nothing at all. They already know what they did. Perfect information on one side, expensive inference on the other, and the expensive side is the side that lost the money.
Seed-commit schemes narrow this, and credit where it is due, they beat nothing. But they still put the burden on the player to record commitments, keep the reveals and check after the fact, and the overwhelming majority never will. A verification only a determined minority can perform is one the house has already priced in.
How verification by proof inverts it
This is the part of Satoshie’s architecture that gets the least attention and deserves the most. When a raffle or a coinflip needs randomness, the Chainlink VRF coordinator does not simply hand a number to the game contract. It verifies the cryptographic proof accompanying that number, on-chain, mechanically, before the callback ever fires. If the proof does not check out, the callback does not happen. There is no human in that rejection decision and no discretion available to anybody, including us.
So the expensive verification has already been performed, by the chain, at the moment of the draw, as a condition of the draw happening. What is left for a player is not an audit. It is a lookup. Open Basescan, find the request, find the fulfilment, read the number that decided your game and confirm it is the same number the contract used. That costs a few minutes and nothing else. No tooling, no sample size, no permission, no arms race.
The gap between those two worlds is not a difference in how honest the operators are. It is a difference in what it costs to find out, and that is the only version of the question a player can actually act on.
The honest limits
Cheap verification of the randomness is not cheap verification of everything, and this series exists because that distinction keeps getting collapsed. The proof covers the number. It does not tell you the front end showed you the right contract, promise the randomness arrives promptly, or audit the inputs a contract chose to trust. Separate properties, separate failure modes, and pretending otherwise is how the industry got here.
One more limit, stated plainly: a proof nobody bothers to check is socially no better than a promise. Cheap verification does not force verification. What it does is remove every excuse and remove the class filter, so checking is no longer a privilege of whoever can rent the best tools this quarter.
The crypto firms writing to the AI labs are asking for parity in an arms race they did not choose. Fine. But the deeper lesson for anyone building games with money in them is that the winning move is not to compete in that race. It is to build systems whose honesty does not depend on the outcome of it. If your fairness needs weapons-grade tooling to confirm, you have not built a fair game. You have built a game that is expensive to catch.
Satoshie runs provably fair raffles and coinflips on Base, with outcomes decided by Chainlink VRF and verifiable by anybody with a block explorer and five minutes.
📷 Photo by Markus Spiske on Unsplash
