For Advisors11 min read

RIA Cybersecurity Requirements When You Build Your Own Tools With AI

A 2025 scan of 1,645 apps built with an AI app builder found about one in ten leaking user data. If you're an advisor building your own intake forms, portals, or automations, here are the security rules that apply to a small firm, plus a checklist you can work through without an IT department.


In March 2025, Matt Palmer, who works in developer relations at Replit, a competing AI coding platform, and a colleague scanned 1,645 web apps listed in the public showcase of a popular AI app builder. They found 170 of them, about one in ten, handing data to anyone who asked: names, phone numbers, payment details, API keys. The apps worked. What was missing or misconfigured was the database rule that says which user can read which rows.

RIA cybersecurity requirements attach to client data wherever it sits, including in software you built yourself. So picture the same mistake in a client portal you put together over a few weekends with an AI coding tool, where the rows are Social Security numbers and tax returns.

If you've moved to an advice-only model, no custodian hands you a client portal. The intake form, the upload link, and the notes system are all things you chose or built, and at a one-person firm nobody else is checking them.

None of this is legal advice. Registration status and state law change the answer, so confirm the details with a compliance consultant or your regulator.

Which RIA Cybersecurity Requirements Apply to a One-Person Firm?

It depends on who you're registered with. A solo firm is probably state-registered, so that column comes first.

State-registeredSEC-registered
Written security policyYes, under the NASAA model where adopted and the FTC ruleYes, in writing since 2005
Written incident response planThe FTC rule excuses firms below 5,000 consumers. Check your stateYes, since June 3, 2026 (December 3, 2025 for advisers with $1.5 billion or more in assets)
Breach noticeState law (every state has one), plus the FTC within 30 days if 500 or more consumers' unencrypted data is involvedAffected individuals within 30 days

State-registered advisers answer to their state's rule and to the FTC. NASAA's Model Rule for Investment Adviser Written Policies and Procedures, adopted November 24, 2020, replaced a 2019 model rule on information security and privacy. It calls for written physical security and cybersecurity policies "tailored to the investment adviser's business model, taking into account the size of the firm." It binds you only if your state adopted it, and we couldn't find a published list of which states have. Search your state securities regulator's rules for "cybersecurity," or ask your examiner.

The FTC Safeguards Rule, by the FTC's own guidance, covers "investment advisors that aren't required to register with the SEC." Along with a written security program, it requires one named person responsible for it (in a solo firm, you), encryption, multi-factor authentication, oversight of service providers, and "secure development practices for in-house developed applications." Firms holding information on fewer than 5,000 consumers are excused from four items: the written risk assessment, penetration testing and continuous monitoring, the written incident response plan, and the annual written report.

SEC-registered advisers fall under Regulation S-P. Amendments adopted in May 2024 added a written incident response program, customer notification, and procedures reasonably designed to ensure a service provider tells you about a breach within 72 hours. The SEC's fiscal 2027 examination priorities list cybersecurity and the protection of customer information among staff's focus areas.

What Have the SEC and NASAA Said About Advisers Using AI?

As of November 2026, the SEC has no rule written specifically for advisers' use of AI. It withdrew its 2023 proposal on predictive data analytics in June 2025.

The enforcement cases to know about involve marketing claims. In March 2024 the SEC settled with two advisers, Delphia and Global Predictions, over statements about AI capabilities they didn't have, with penalties of $225,000 and $175,000. If your website calls your planning "AI-powered," be ready to show what that means.

On the state side, a NASAA project group's compliance article on AI, which isn't formal NASAA policy, tells advisers to vet a tool before entering sensitive client data, review AI output before it reaches clients, and document that review. NASAA's 2026 report says guidance on AI is being drafted.

Then there are records. For SEC-registered advisers, Rule 204-2 requires keeping written communications about advice and recommendations for at least five years. A Cooley analysis from April 2026 suggests asking whether an AI note taker's transcript is the only or most complete record of something the rule requires you to document, and notes that one sent to a client becomes a written communication you have to retain. State recordkeeping rules vary, so check yours.

What Counts as Nonpublic Client Information?

Under Regulation S-P, personally identifiable financial information includes anything a person gives you to obtain your services, account balances, and "the fact that an individual is or has been one of your customers." Those definitions bind SEC-registered firms, and the FTC rule's are nearly word for word the same. So your client list counts, and so does what a prospect types into an intake form linked from your marketing.

Both rules leave out information that doesn't identify a person, such as "blind data that does not contain personal identifiers such as account numbers, names, or addresses." So before you type a client's situation into a chatbot, take out the name, employer, town, and account numbers, and round the dollar figures. "A 58-year-old teacher with about $900,000 in a 403(b) and a pension starting at 62" gets you the same analysis. That's a practical reading, not a safe harbor, and an unusual set of facts can still point to one person.

That doesn't work for an uploaded statement, tax return, or transcript, which carry names and account numbers on their face. Redact the file first, with Social Security numbers blacked out, or upload it only to a tool that has passed the checks below. And don't collect clients' brokerage passwords. Ask for statements or view-only access.

Can Financial Advisors Put Client Data in ChatGPT or Another AI Tool?

That depends on the plan. OpenAI's published terms, for example, say it may use content from its individual plans to train models unless you opt out, and that it doesn't train on business plan or API data by default. Terms differ by vendor and change over time, so read the current ones for the plan you're on.

For any AI or cloud tool that will hold client information, including the AI coding tool you build with, find out:

  • Is my content used to train models, and is that a default or a setting?
  • How long is data kept, and can I delete it, including from backups?
  • Where is it stored, and which subcontractors can reach it?
  • Will the vendor sign a data processing agreement, and how fast does it promise to report a breach?
  • Is there an independent audit I can read, such as a SOC 2 report, and does its scope cover the product I'm using?
  • Does it support multi-factor authentication, and can I export my records?

Here's how that goes for a hypothetical AI note taker. The vendor and its answers are invented.

The terms say paid plans exclude customer content from model training, while the free plan allows use "to improve services." So it's the paid plan or nothing. The data processing agreement promises breach notice "without undue delay," with no number attached, so if you're SEC-registered you email the vendor asking it to commit to 72 hours and save the reply. A SOC 2 report is available under a nondisclosure agreement, so you request it and read the exceptions. Multi-factor authentication is off by default, so you turn it on.

The decision: use the paid plan, export each meeting summary to the client's file, and get each client's consent before you record. About 11 states require consent from everyone on the call, and with a client in another state the stricter law may apply. Then write half a page with the date, what you checked, and the links, and add the tool to the vendor list in your security policy. Expect an hour or two of your own time, spread over the days the vendor takes to answer.

A Security Checklist for a Client Portal or Intake Form You Built Yourself

When Veracode tested code from more than 100 AI models in 2025, 45% of the samples failed security tests.

You can do most of the checklist below yourself, from the settings pages of tools you already use. Effort figures are rough estimates for a solo firm.

ControlWhyEffort
Multi-factor authentication on email, hosting, database, and AI accountsCISA says users who enable it are "significantly less likely to get hacked"30 minutes
Secret keys and passwords kept on the server, never in code or the browserAnything sent to a browser can be read there. Keys your provider labels public are meant to be seen, so the next row matters1 hour
Access rules on every database table, so each client reaches only their own rowsThis is the gap behind the 170 leaking apps1 to 2 hours
File storage private by default, shared through expiring linksA public link to a tax return works for anyone who has it30 minutes
Encryption in transit and at rest, including your laptop's diskThe FTC rule requires it for firms it covers15 minutes to confirm
Automatic backups, plus one test restoreCISA calls testing restores "critical"1 hour
Access logs kept, a separate admin login, and keys limited to one job eachYou can tell what was taken, and one stolen credential reaches less1 hour
Fake data until the app passes testing, then an outside reviewMistakes surface before real clients are in the system, and real records stay out of the coding toolA few developer hours
Updates applied monthly, with your host's security alerts onNobody else is patching an app you built30 minutes a month
No Social Security numbers unless a task needs themYou rarely need one if you don't open accounts or file returnsNone

If an outside review isn't in the budget, buy a vetted tool for the parts that hold client documents and build only the parts that don't.

An Incident Response Plan That Fits on One Page

SEC-registered firms must have a written incident response program, and the NASAA model's policies have to cover response and recovery. One page can hold it:

  • Who. You decide, with a backup person who knows where this page is. List your vendors' security contacts, your compliance consultant, an attorney, and your cyber insurer.
  • Contain. Revoke keys, reset passwords, take the app offline. The FTC's breach response guide warns against destroying forensic evidence, so export the logs before you wipe anything.
  • Assess. Which clients, which data, over what period.
  • Notify. Write your deadlines here now, from the breach notice row in the first table and the breach laws of your state and your clients' states.
  • Record. Write down what happened and what you decided, including a decision not to notify. For SEC-registered advisers, that's a required record under Rule 204-2.

Print it, so you can read it when your systems are down.

If you've already built something, test it for the failure behind those 170 apps. In a private browser window, logged out, try to load a client record directly. Then sign in as one test client and try to open another's. Passing both doesn't clear you. The scan found the data through requests sent straight to the database, using a public key built into the page, so an app can hide records on screen while its database still returns them. Open your browser's developer tools, find a request your app makes to its database, and resend it without signing in, or have your outside reviewer do it. If rows come back that shouldn't, take the app offline until it's fixed.

Have a lesson about building or securing your own advisor tech? Share a proposed contribution with the Advice-Only Network at [email protected], using the subject "AON Blog Post Submission."