Who Pays When An AI Trading Agent Messes Up

The Lobstar Wilde mistake and why it matters

On February 22, the AI trading bot Lobstar Wilde tried to send a small crypto tip and, instead, moved roughly a quarter of a million dollars out of its own wallet. The bot had been spun up days earlier by an OpenAI employee with 50,000 dollars and told to "make no mistakes," yet it stumbled within three days of going live.

It was not linked to any brokerage account, so nobody's nest egg was exposed. The misfire shows how an automated agent can misread a simple instruction and fire off an action before anyone can step in.

That is precisely the kind of hazard regulators and Congress have been examining while Robinhood, Public, SoFi, Coinbase and Webull introduced AI trading agents for retail customers this year.

What the platforms' disclosures say about who bears risk

Finder dug into each platform's formal disclosures, not the marketing pages, and the pattern is clear: investment decisions made by an agent sit on the customer's shoulders.

• Robinhood Agentic Trading: Customers accept all risk for trades an agent executes, and Robinhood says it does not control, supervise or audit third party agents. Guardrails include a separate, dedicated account, a push alert for each trade, support that can match your instruction to what the agent did, and the ability to disconnect at any time.

• Public Agents: Customers alone judge whether a strategy fits and must verify instructions before turning an agent on. Actions happen inside one authenticated brokerage environment, and each agent needs explicit approval before it goes live.

• SoFi Composer: The retail investor makes the final call on every trade. Composer says it does not provide recommendations or monitor specific investments unless it says so in writing, notes any departure from an AI strategy's logic is at the client's own risk, and reminds that results are not guarantees. Customers can predefine trading rules and backtest those rules using historical data.

• Coinbase for Agents: It warns outputs may be wrong or partial and says they are not investment advice, so customers are fully responsible for their trading. Guardrails include an isolated sub account plus user defined limits on capital, allowed assets and maximum order size.

• Coinbase Advisor (separate product): It is registered as a fiduciary and says it still cannot "replace your judgment for your trades," with customers responsible for their activity. The firm is registered as an investment adviser with the SEC and as a commodity trading adviser with the National Futures Association, and it will not act without the user's prior approval.

• Webull MCP Server: Customers are responsible for verifying orders and monitoring positions, and Webull disclaims liability for losses tied to automated or AI directed decisions. The service ties directly into the main brokerage account, and Webull's own disclosures do not describe a separate sandboxed sub account.

Across the set, firms use different guardrails, but the liability consistently sits with the user.

Smart investors look for tools that help reduce mistakes and manage risk. We're hosting our first ever investor livestream on 11/17 to show you how to use ChatGPT and Claude to invest like a pro, hosted by Briefs Finance CEO Jaspreet Singh. Save your seat.

Where SIPC fits - and where it does not

SIPC is the nonprofit backstop for brokerage failures. If a broker collapses and customer assets are missing, SIPC works to restore them, up to 500,000 dollars per customer, with a 250,000 dollar ceiling on cash.

What it does not handle is investment outcomes. If an AI agent in a solvent account buys something that plunges, or executes a trade you did not intend, that is not a missing assets problem. It is an investment decision issue, and SIPC was never meant to cover it.

Regulators are already flagging the gap

FINRA, which oversees broker dealers, called out autonomous AI agents as a leading concern in its 2026 regulatory report. It warned that a poorly designed reward function could push an agent to optimize in ways that hurt investors.

FINRA says firms need a supervision, governance or model risk framework around these tools rather than relying on the tech to police itself. Its rules are technology neutral, so they cover AI the same way they cover any other tool a firm uses, but they were not crafted with stand alone, third party agents in mind.

Lawmakers see the same hole. In June, the House Financial Services Committee asked the SEC to clarify who is accountable when an agent, not a human, makes the call and whether a broker dealer's duties persist when a third party AI is placing trades. As of now, no public SEC response has been reported.

The practical takeaway for your portfolio

The protection that exists today was built for broker failures, not agent misfires. In plain English: your real safeguards are how your account is set up and the limits you turn on. A separate pot of money you can afford to lose, tighter spending and symbol limits where offered, manual approvals if available, and frequent checks of your activity feed all reduce blast radius.

Even so, those measures would not have prevented Lobstar Wilde, which was not a customer of these platforms and executed too fast for any human to step in. That is the risk investors accept with agentic trading, whether it is a weekend project or wired into a brokerage account. Keep a close read on your platform's liability language and the controls it provides - those details decide who absorbs the loss when an AI agent gets it wrong.

Learning how new tools change decision making can give investors a subtle edge. On 11/17, our CEO Jaspreet Singh goes live for our first ever investor livestream on how to use ChatGPT and Claude to invest like a pro. Register to join him.