September 30, 2026

An AI Can Trade Your Stocks Now. You Carry the Risk.

Robinhood is putting a trading agent inside its app, and the fine print puts every loss on the customer. On the same day, Washington chose voluntary promises for the AI labs.

By Jay Cadmus

A phone lies face-down beside coffee and a handwritten page on a kitchen table in early morning light.

The Short Version

On Tuesday in Houston, Robinhood showed its most active customers a new feature. You name an AI agent, move some cash into a separate account and pick a model from OpenAI or Anthropic. Then the agent can research stocks and place trades for you around the clock. More than 150,000 customers have already opened accounts for outside agents, and the company is now building one into the app for everyone. The terms are blunt: you assume all risk for your agent's orders, and Robinhood says it doesn't supervise or audit agents. Earlier that day at the White House, six AI companies signed a voluntary accord promising internal controls and outside audits of their most powerful models. President Trump called it "morally binding." The two events came from different worlds on the same day. Together they show who is setting the rules for AI that touches your money right now: mostly the companies selling it, with you as the backstop. Below is what Robinhood turned on, what actually protects you, and how I'd test an agent with money I could afford to lose.

What Robinhood Turned On

The product is called Robinhood Agents, and setup takes about four steps. You name the agent, open a dedicated agent account, choose a model, and set a strategy and limits. The agent can only spend what sits in that dedicated account. Approval of each trade is switched on by default, and you can turn it off at any time. Robinhood is also building a feature called Loops, which turns a strategy into a standing instruction. A Loop might check the market every morning and buy when a condition is met, or run a strategy overnight while you sleep.

Timing depends on who you ask. Robinhood's own announcement says the agents are coming soon to eligible U.S. customers, while Fortune reports that access for the company's roughly 29 million customers starts this week. Yahoo Finance describes a randomized rollout, with attendees at the Houston event getting access first. In a demonstration Fortune watched, users could choose OpenAI's GPT-6 Luna or GPT-6 Sol, or Anthropic's Opus 4.8. Luna is free until the end of the year, and the other models are billed at standard usage rates.

This builds on something Robinhood opened in May for technically inclined customers, who could connect their own agents to their accounts. The company says more than 150,000 of them have opened agent accounts since then, and that agents now call Robinhood's tools almost 30 million times a day. Those are the company's figures, and they measure activity, which is different from results. Abhishek Fatehpuria, a Robinhood product vice president, told Yahoo Finance that making the setup far easier could take users from 150,000 to millions.

Then there's the fine print. Robinhood's agent page warns that agents can make errors and misread instructions. It puts the burden in one sentence: "You assume all risk for orders placed by your AI agent for execution and for any use of your data by third-party AI providers." Anything you share with the AI provider also leaves Robinhood's security environment and falls under that provider's terms.

The summit announcement adds a line every user should read twice: "Robinhood does not control, supervise, monitor, recommend, or audit agents." The company also told Yahoo Finance it hasn't yet compared the investment results of agent accounts with customers' other strategies.

My read is that this is an honest disclosure. It tells you plainly where the weight sits, and it sits with you.

Who Writes the Rules for an Agent With Your Money

Robinhood's legal position, as Fortune describes it, is that hosting agents is different from giving financial advice. The company likens an agent's suggestions to a customer asking the internet or a friend. Fortune also notes that nobody has settled where legal liability falls if an agent trade goes wrong.

The industry's own regulator saw this coming. FINRA, which oversees brokerage firms, listed AI agents as an emerging risk in its 2026 oversight report. Its concerns read like a preview of this week. Agents might act without a human checking their work, or reason through so many steps that their choices are hard to trace afterward. They "may act beyond the user's actual or intended scope and authority." For member firms, the report suggests keeping a human in the loop and tracking what agents do.

Washington's answer to AI risk arrived the same day, aimed at a different layer. At a White House lunch on Tuesday, Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the White House Accord on Super Intelligence. It commits them to four layers of control. They'll monitor their models during training and use, and an internal team will check those controls. Independent auditors will test them, and independent board committees will review the audit reports. Speaker Mike Johnson called it a statement of principles. Al Jazeera reported that it sets no deadlines and no enforcement mechanism. Vice President JD Vance put the administration's view simply: "The solution to some of the AI risks is for you guys to take the risk seriously, not to come to the government for a regulatory regime."

Critics were quick. Toby Walsh of the University of New South Wales asked, "What other trillion-dollar industry marks its own homework?" In the Senate, Mark Warner tried to pass a bill creating an AI Safety Board with enforceable standards. It would get access to new models 45 days before release. Ted Cruz objected. Trump also signed an executive order telling federal agencies to call AI "super intelligence," with proposed legislative language defining the term due within 60 days.

Put the two stories side by side and the gap is easy to see. The accord's commitments are written for companies that train and deploy frontier models. Your Robinhood agent may run on a signatory's model. Still, the choices that decide your outcome are set by the brokerage's product and by you: whether approvals stay on, how big a trade can be, and who pays for a mistake.

When Thousands of Agents Move at Once

The risk also reaches past any one account. On Sunday, Apollo's chief economist Torsten Slok warned about what he called an agentic bank run. Checking accounts pay about 0.1 percent on average, he noted, while some fintech accounts pay 3.3 to 5 percent. AI assistants could soon sweep household cash toward the higher rate automatically. "If every household used AI agents to optimize the return on their cash balances, banks could lose a large share of the cheap deposits they rely on to make loans," he warned.

CoinDesk raised the trading version of the same worry. Thousands of agents running similar models on similar data could pile into the same trade at the same moment and turn a small market move into a sharp one. For now that risk is largely theoretical, because the technology is so new. Your agent's "independent" judgment may look a lot like everyone else's.

OpenAI offered a different design at its developer conference, also on Tuesday. Its new Decisions API limits a model to a set of questions the customer defines, each with a fixed list of possible answers, such as approve or reject. That narrows what the AI can do, which makes its choices easier to check. I'd judge any money-moving agent by the same standard: how small is the set of things it's allowed to do, and can you see every one of them afterward?

A Small Test With Money You Can Lose

Take Dana, an electrical contractor with a brokerage account that mostly holds an index fund. She likes the idea of an agent that buys a few favorite stocks on bad days, and she's curious enough to try it.

Before she moves a dollar, she writes one page of rules for the agent. It lists which stocks it can buy and the most it can spend on a single trade, and it bans options, margin and crypto. She keeps a copy, because the agent's instructions and her intentions need to match, and FINRA's warning about agents acting beyond their intended scope is exactly that failure. Then she sets a bounded pilot. She moves $1,000, an amount she could lose without changing her plans, into the agent account and leaves trade approvals on for the first 60 days.

Her baseline is simple. The same $1,000 stays in her index fund on paper, and at the end she compares the two after every cost, including model usage fees once the free period ends. Every trade the agent proposes goes in a log, along with whether she approved it and how it turned out, so a rejected idea counts as the agent's miss. Her own time goes in the log too, because reading proposals every morning is real work that's easy to forget.

She knows where it can break. The agent can misread an instruction, act on stale information, or chase the same trade as thousands of other agents. So she decides in advance who stops it and when. She does. One trade outside her written rules, or a loss of 15 percent against the baseline, and she turns the agent off and moves the cash back. If it beats the index fund after costs and her time, she earns the right to consider a second round with the same limits.

Opportunity Radar

One small business idea stands out. Robinhood says it doesn't monitor agents, and its terms tell customers to review their own account activity. For a busy person, that's a chore that slips. The offer is a weekly plain-English report that checks each agent trade against the customer's written rules and compares results with a simple baseline, like an index fund. Active traders and fee-only financial advisers with clients who use agents are the likely buyers. Test it by hand with ten people who export their statements, at $15 a month. If fewer than three renew after the first month, or Robinhood ships its own version, walk away.

What You Can Do With This

If you invest on your own

Leave trade approvals on until you've watched the agent for at least a couple of months. Fund the agent account with an amount you'd be comfortable losing, and write your rules down before you start. Read the risk disclosure in full, since it spells out that the losses are yours.

If you run a small business

Look at what your operating cash earns today and decide whether you'd let software move it. If you do, set a hard limit on how much it can move and where it can go. Slok's warning is about households acting all at once, and your own cash decisions still deserve a human signature.

If you advise clients

Ask clients directly whether they've connected an agent to any account. FINRA's list of agent risks makes a solid intake checklist: scope, human approval, and whether you can trace what happened afterward. A client's agent account can undo the plan you built together.

If you build agents

Give your agent the smallest set of actions that does the job, and log every one. Fixed choices, like the approve or reject design OpenAI previewed this week, are easier for customers to trust and easier for you to defend.

The Bigger Picture

In one day, the question of who answers for AI that acts on its own got two answers, and both came from promises and contracts. The labs promised to police their own models, and a brokerage told its customers that the risk of an agent's trades is theirs. That puts terms of service ahead of regulators as the working rulebook for AI that handles money. When the first large agent-driven loss happens, the fine print customers clicked through will decide much of who pays, and courts or regulators will likely be catching up. Until that changes, the protection that matters most is the one you set yourself: a small account, approvals on, written rules, and a clear point where you pull the plug.

References