Binance 现允许 AI Agent 交易，但其约束主要仍由用户负责

Binance 是全球最大的加密货币交易所，注册用户超过 3 亿。该公司于周四推出了一个允许 AI Agent 代表用户分析市场并执行交易的平台，将自主 AI 直接带入真实资金管理领域。

这个名为 Agent OS 的平台允许开发者将 AI 应用和 Agent 连接至 Binance 的金融基础设施。它整合了该交易所既有的工具和服务，包括 Binance API、Binance Wallet Agentic Hub、Binance x402 交易验证与支付促进 API，以及 Binance Skill Hub；同时还新增支持其 模型上下文协议（Model Context Protocol，MCP）。该平台也可与 OpenAI 的 ChatGPT 和 Codex、Anthropic 的 Claude Code 以及 Cursor 等工具配合使用，用户可以授权 Agent 访问市场数据、查看账户信息并执行交易。

Binance 的 Agent OS 允许 Agent 代表用户交易
图片来源： Binance

然而，随着 AI 竞赛从回答问题的聊天机器人转向能够采取行动的 Agent，Binance 将监督和约束 Agent 的大部分责任交给了用户。用户最终必须决定 Agent 可以访问什么、可以交易什么，以及为其操作设定哪些限制。

Binance 产品副总裁 Jeff Li 在接受采访时表示：“我们不会给予完全自由，而是将权力交到用户手中，让他们能够细粒度地控制 Agent 可以做什么。”他说：“我们将[控制权]置于账户层面，以保护用户资金。”

Binance 主要通过专门的“子账户”实现这一点。用户可以将子账户分配给 Agent，并为特定活动配置账户，例如现货或期货交易。Li 告诉 TechCrunch，这些子账户默认禁止提现，从而在 Agent 的活动范围外建立了一个沙箱。

一位 Binance 代表表示，用户还可以选择要求 AI Agent 在每笔订单执行前都寻求批准，或者在权限配置完成后让其自主执行交易。Binance 并未对 AI Agent 的交易金额或亏损金额设置单独上限，因此，用户转入子账户的资金数额实际上便构成了限制。

Binance 的子账户可以设置特定交易限额和权限
图片来源： Binance

当被问及 Binance 是否能够了解 Agent 做出某一特定交易决策的原因时，Li 表示，推理过程发生在 Binance 系统之外，要么在用户的电脑上，要么在用户选择的 AI 应用内部。他说：“我们实际上无法看到用户操作背后的推理过程。”

这意味着 Binance 可以监控 Agent 最终产生的交易活动，但对于某项决策是否受到错误信息或操纵影响，其可见性有限。

当被问及如果 Agent 遭到提示注入攻击操纵，或以其他方式被攻破，将会发生什么时，Li 再次指出子账户是主要防线。Binance 还表示，其现有针对子账户 API 的安全、风险控制和反洗钱政策将在 Agent OS 推出时适用。

交易是 Binance 瞄准的首批应用场景之一。不过，Li 表示，Agent 还可以监控市场、开展研究和风险分析、响应信号，并自主下单或执行套利等策略。

Agent OS 的设计还旨在让 Agent 连接支付和链上活动。通过 Binance 的 x402 集成，Agent 可以发起并结算支付；其 Agentic Wallet 则允许 Agent 与代币及去中心化金融协议互动。

与交易所交易不同，Binance 不会对 Agent 在子账户内可交易或可亏损的金额设置单独上限；但 Agentic Wallet 交易设有 Binance 规定的每日限额。该公司称，常规兑换的每日上限为 5 万美元，DeFi 交易默认每日限额为 10 万美元，x402 支付则被限制为每日 20 美元。

Li 表示，Agent OS 是 Binance 向开发者提供平台、让其构建可在加密货币和传统市场中行动的 AI 驱动应用的“第一步”。

Binance 并非唯一向 AI Agent 开放基础设施的公司。竞争对手加密货币交易所也在朝同一方向发展，利用 MCP 及其他开发者工具，让 AI 应用直接访问市场数据和交易系统。

今年 3 月，Kraken 推出了一款内置 MCP 服务器的开源命令行工具，允许 AI Agent 执行包括现货和期货交易在内的操作。Coinbase 随后于 6 月推出 Coinbase for Agents，将 AI Agent 直接连接至用户账户，并允许它们在用户设定的限制内进行交易、付款及执行其他金融工作流。同样，OKX 今年早些时候通过引入一个开源 MCP 工具包，在其平台上启用了 Agent 交易。

Binance, the world’s largest crypto exchange with more than 300 million registered users, on Thursday launched a platform that lets AI agents analyze markets and execute trades on users’ behalf, bringing autonomous AI directly into the business of managing real money.

Called Agent OS, the platform lets developers connect AI applications and agents to Binance’s financial infrastructure. It brings the exchange’s existing tools and services such as Binance APIs, Binance Wallet Agentic Hub, Binance x402 transaction verification and payment facilitator API, and Binance Skill Hub, along with newly introduced support for its Model Context Protocol (MCP). The platform also works with tools including OpenAI’s ChatGPT and Codex, Anthropic’s Claude Code, and Cursor, allowing users to authorize agents to access market data, view account information, and execute trades.

Binance’s Agent OS lets agents trade on users’ behalfImage Credits: Binance

However, as the AI race moves away from chatbots that answer questions to agents capable of taking action, Binance is putting much of the responsibility for keeping them in check on users, who ultimately have to decide what agents can access and trade and set limits on what they can do.

“Instead of total freedom, we put the power in users’ hands to give them the granular access control of what they can do through the agent,” said Jeff Li, vice president of product at Binance, in an interview. “We put [the control] at the account level to protect the users’ funds.”

Binance does that primarily through dedicated “sub-accounts”, which users can assign to agents and configure for specific activities, such as spot or futures trading. Withdrawals from those sub-accounts are blocked by default, Li told TechCrunch, creating a sandbox around an agent’s activity.

Users can also choose whether an AI agent must seek approval for every order or can execute trades autonomously once its permissions are configured, a Binance representative said. Binance does not impose a separate cap on how much an AI agent can trade or lose, so the amount a user transfers into the sub-account effectively serves as the limit.

Binance’s sub-accounts can have specific transaction limits and permissionsImage Credits: Binance

Asked whether Binance can see what leads an agent to make a particular trade, Li said the reasoning happens outside its systems, either on the user’s computer or within their chosen AI application. “We really cannot see the reasoning of what the user’s action is,” he said.

That means Binance can monitor an agent’s resulting trading activity, but has limited visibility into whether a decision was influenced by faulty information or manipulation.

Li again pointed to the sub-account as the main line of defense when asked what would happen if an agent were manipulated through a prompt-injection attack or otherwise compromised. Binance also said its existing security, risk-control, and anti-money-laundering policies for subaccount APIs apply to Agent OS at launch.

Trading is one of the first use cases Binance is targeting. Nonetheless, Li said agents could monitor markets, conduct research and risk analysis, react to signals, and autonomously place orders or execute strategies such as arbitrage.

Agent OS is also designed to connect agents to payments and on-chain activity. Through Binance’s x402 integration, agents can send and settle payments, while its Agentic Wallet allows them to interact with tokens and decentralized-finance protocols.

Unlike exchange trading, where Binance does not impose a separate cap on how much an agent can trade or lose within its subaccount, Agentic Wallet transactions carry Binance-set daily limits. Regular swaps are capped at $50,000 a day, DeFi transactions have a default $100,000 daily limit, and x402 payments are limited to $20 a day, according to the company.

Li said Agent OS was Binance’s “first step” toward giving developers a platform to build AI-powered applications that can act across crypto and traditional markets.

Binance is not alone in opening its infrastructure to AI agents. Rival crypto exchanges have been moving in the same direction, using MCP and other developer tools to give AI applications direct access to market data and trading systems.

In March, Kraken launched an open-source command-line tool with a built-in MCP server that allows AI agents to execute actions including spot and futures trades. Coinbase followed in June with Coinbase for Agents, which connects AI agents directly to users’ accounts and allows them to trade, make payments and execute other financial workflows within user-set limits. Similarly, OKX enabled agentic trading on its platform by bringing an open-source MCP toolkit earlier this year.