Six of ten AI models paid twice for the same data in a controlled Solana benchmark. Nata’s wallet checklist points to a check that spending caps alone can’t provide.

Nata, a crypto commentator who posts as @HUNHOfamilyTH, used a recent Solana benchmark to make a case for enforcing AI spending rules in the wallet or smart contract. In a November 6 post, Nata proposed spending caps, a list of approved recipients and an emergency stop. One result gives those ideas a concrete test: six of the ten models paid twice for data they had already bought.

这三道都写进钱包或合约。

— Nata (@HUNHOfamilyTH) November 6, 2026

What the Solana test measured

BlockRun’s Agent Money Benchmark used real USDC payments on Solana mainnet in a controlled test from October 27 to 29. Each model ran 76 financial research tasks and 80 marketplace tasks, where sellers were deliberately set up to try tactics such as asking for a second payment. Seven models paid a cheating seller at least once.

According to the published methodology, the agents were told they had a budget but received no specific payment rules. Each model attempted each task once. The results show which mistakes occurred in that setup. They leave open how often the same mistakes would occur with different instructions or wallet controls.

The missing check is whether the order was already paid

Nata’s caps would limit how much a mistake can cost. An approved recipient list would narrow where the money can go. A second charge from an approved seller can still fit inside both limits. The payment software also needs to know which order the charge belongs to.

For a team building an agent that buys market data, that means giving each purchase an order ID and checking its payment status before authorizing another transfer. If a seller’s response goes missing, the software can first check whether the original transfer settled. The practical test is whether the agent can recover from a failed response without paying for the same order twice.