Binance now lets AI agents trade crypto for you — but the guardrails are on you

What just launched

Binance, the world's largest cryptocurrency exchange with more than 300 million registered users, has launched a new platform called Agent OS. The idea is simple to describe and, in practice, a little unsettling: it lets AI tools — including OpenAI's ChatGPT and Codex, Anthropic's Claude Code, and the coding tool Cursor — watch crypto markets and automatically place trades on someone's behalf, without a human needing to click "buy" or "sell" every time.

How Binance is trying to keep this safe

Rather than simply handing an AI agent full control over your main account, Binance requires you to set up a separate "subaccount" specifically for agent-based trading. Money you transfer into that subaccount is walled off from the rest of your holdings, and withdrawals from it are disabled by default — so even if an AI agent made every possible mistake, it couldn't directly drain your main wallet. You also decide for yourself whether the agent needs your approval for every individual trade, or whether it can operate independently within limits you set.

There are no separate trading limits beyond the amount you choose to deposit into that subaccount yourself — your own decision about how much to put in is, in effect, your safety net. For actual money transfers and payments (as opposed to trades), Binance does apply its own separate daily limits, such as a $50,000 cap on regular transfers.

The part Binance can't fully see

Here's the catch: Binance can see which trades an AI agent actually executes, but not the reasoning behind them, because that thought process happens outside Binance's systems — on your own computer or within whatever AI app you're using. That means if an agent is fed bad information, or manipulated in some way, Binance would see the resulting trade but not necessarily know why it happened.

Not the only exchange doing this

Binance isn't alone here — rival exchanges Kraken, Coinbase, and OKX have all rolled out their own versions of AI agent trading in recent months, each taking a similar approach: give the AI real trading power, but keep the safety boundaries in the hands of the user rather than fully automating trust.