AI agents can be deceived, and this issue won’t disappear just by switching to a smarter model. BNB Agent Studio, in collaboration with AWS, conducted a series of “deceptive agent” tests. Attackers inserted incorrect recipients or amounts into otherwise normal tasks; in some scenarios, up to 87% of the six tested models executed unauthorized payments. Even with a wallet policy layer added, the models still attempted to make errors—but in the test environment, not a single unauthorized payment was signed.收款白名单、消费上限和签名权限由代码执行，私钥也不进入模型上下文。 I agree with this approach. AI can help you make decisions, but spending boundaries must be hard-coded into the wallet. Current results are based on single-round tests and BSC testnet validation; there’s still a long way to go before achieving long-term production-grade protection. But the direction is clear: the defense must be placed before signing.