Bybit Launches AI Sub-Accounts to Ringfence Agent Trading Risk

• Bybit has launched AI Sub-Accounts to give traders safer ways to connect AI agents to live trading accounts.
• The new account type ringfences agent activity with asset caps, transfer limits, leverage controls and API-only execution.

Bybit is putting a security wall around AI-driven trading. The Dubai-based exchange has launched AI Sub-Accounts, a dedicated account structure designed to let users connect AI agents without giving them broad access to an entire portfolio.

AI agents get isolated trading environments

The new account type is separate from Bybit’s regular, custodial and Islamic sub-accounts. It is now live for all users and becomes the default structure for traders connecting AI agents to the platform.

The reason is fairly straightforward. As AI agents move deeper into trading, they also create a new class of account risk. A poorly configured bot, compromised model, vulnerable codebase or rogue agent could place unauthorized trades, trigger liquidations or move too aggressively through a user’s balance. In crypto markets, where positions can be leveraged and settlement is fast, that can become expensive very quickly.

Bybit’s AI Sub-Account is built to contain that risk. Authorized agents operate inside a ringfenced account environment, separated from the user’s primary funds and other sub-accounts. Fund movement is contained within the designated AI account, while users can set boundaries around asset holdings, transfer limits, leverage caps and permitted commands.

Permission controls become part of AI trading

The design also removes login access for agents. Execution happens through APIs only, reducing the risk of account hijacking or manual access to AI-controlled funds. Parent accounts retain read-only oversight, giving users visibility into agent activity without needing to intervene constantly.

That matters for traders testing new strategies. Users can assign experimental AI agents to isolated sub-accounts, monitor performance and only expand use once the behavior looks stable. It is less dramatic than letting an agent trade freely, but probably more realistic.

Wu said the setup prevents AI agents from controlling an entire account or moving assets unpredictably.

The launch points to a broader shift. Exchanges are no longer only building tools for human traders clicking through screens. They are preparing for software agents that act continuously, react quickly and need tighter limits than humans usually set for themselves.