Binance lets AI agents trade real money on its exchange
Two moves today show the AI-agent economy maturing fast — one putting autonomous software directly in charge of real money, the other raising hard questions about how real China's robotics boom actually is.
Binance has launched Agent OS, a platform that lets AI agents analyze markets and execute trades on users' behalf using real money. The world's largest crypto exchange, with more than 300 million registered users, opened the platform Thursday, wiring agents into Binance's APIs, wallet infrastructure, and payment rails and giving them access to tools including ChatGPT, Claude Code, and Cursor. Developers connect agents through a Model Context Protocol server, so an agent can pull market data, view an account, and place orders once a user grants permission.
The safety model is where it gets interesting — and thin. Binance leans on dedicated sub-accounts that users assign to agents, with withdrawals blocked by default and permissions scoped to specific activities like spot or futures trading. Users choose whether an agent must ask before every order or can trade autonomously, but Binance imposes no separate cap on how much an agent can trade or lose; the balance you fund into the sub-account is the limit. Crucially, Binance says it cannot see the reasoning behind an agent's decisions — that happens on the user's machine or inside their AI app — so if an agent is manipulated through a prompt-injection attack, the sub-account sandbox is the main line of defense. That is a real gap for software that can be talked into bad trades by a malicious webpage. Rivals are moving the same way: Kraken shipped an open-source MCP command-line tool in March, Coinbase followed in June with "Coinbase for Agents," and OKX enabled agentic trading earlier this year.
An FT investigation suggests China's humanoid-robot demand is propped up by a circular, state-backed scheme rather than genuine commercial uptake. According to reporting summarized by The Decoder, Chinese humanoid makers sell a large share of their machines to state-backed training centers, where people teleoperate the robots through physical tasks and then sell the collected motion data back to the manufacturers — a loop that blurs the line between real demand and policy-created demand. At Unitree, nearly three-quarters of humanoid revenue in the first nine months of 2025 came from education and research, and analysts question both the valuation (about 35.89 times revenue versus roughly 20 times for its Hong Kong rivals) and the data's usefulness, since only two to three of every eight training hours are said to be usable.
The pattern echoes the circular-investment criticism leveled at Nvidia in the US, where backing customers who then buy its products inflates the apparent size of the market. We dug into Unitree's blockbuster debut and the broader China robotics land-grab earlier this week — Unitree's 629% debut and the hunt for China's next robot champion — and the FT's reporting is the skeptical counterweight to those eye-popping numbers.
What to watch: whether Beijing's continued subsidies turn this into a real industry (as with solar and EVs) or leave investors holding overvalued metal.
If an AI agent you authorized lost your money to a prompt-injection attack, who should be on the hook — the exchange, the agent's developer, or you? Tell us in the comments.
