Most office security training still runs on the lesson it has taught for years. Look for spelling mistakes. Be suspicious of "Dear Customer." Hover over the link before you click. That advice worked when phishing was written by people in a hurry. It works far less well now that attackers use AI tools that write clean, personal email in seconds, copy the way a colleague writes, and clone a familiar voice.

Our Threat Intelligence Center analysis covers a roundtable of editors from Dark Reading, TechTarget and Cybersecurity Dive on what has to change. One figure from it stands out. SANS found that 19% of organizations run awareness training mainly to meet a compliance or audit requirement. In a small office, that usually means a video once a year, a short quiz and a completion certificate in a folder. The certificate satisfies the checklist. It does little to prepare the receptionist for a call that sounds exactly like the office manager.

What the Warning Signs Look Like Now

The best-known case is Arup, the British engineering firm. In early 2024 an employee in its Hong Kong office joined a video call with people who looked and sounded like the company's chief financial officer and several colleagues. Every one of them was a deepfake. The employee made 15 transfers totaling about $25 million. The fraud came to light only when the employee later checked with head office.

That last detail is the lesson. The employee did eventually make the right move, a check through a separate channel. It simply came after the money had left.

Three other changes matter for a practice or firm:

• Help desk resets. Attackers call an IT help desk pretending to be an employee, sometimes with a cloned voice, and ask for a password or multi-factor reset. If the reset goes through, they hold a working account, and everything they do with it looks like normal use.
• Fake AI tool downloads. Staff looking for a new AI assistant can land on a sponsored search result or a social media ad for a fake download. Researchers at Gen Digital found that nearly a third of the Facebook and Instagram ads they examined pointed to a scam, phishing or malware. The fake pages are often hosted on well-known platforms, so checking the web address offers little protection. The download installs malware that copies saved passwords and logged-in sessions.
• Client data pasted into free AI tools. An employee who pastes a patient note, a contract or a client's tax figures into a free chatbot has handed that information to a service your office has no agreement with and no record of.

Teach the Request, Not the Red Flags

The fix the roundtable keeps returning to is a change in what you ask staff to judge. Deciding whether a message looks fake is a test people will now fail, because a good fake looks exactly right. What staff can control is how they confirm certain requests before acting on them.

For most practices and firms, four kinds of request cover it:

1. Moving money, including any wire or payment outside the normal schedule.
2. Changing payment details, such as a vendor's new bank account.
3. Resetting a password or multi-factor sign-in.
4. Sending sensitive records, such as patient files, client documents or payroll data.

Each of those gets confirmed by calling the person back on a number you already have, from your own contact list or vendor file. Never use the phone number, email address or meeting link that came with the request. A cloned voice cannot answer a call placed to the real person's phone. Write the rule down, give it to everyone who handles money, records or accounts, and apply it to the owner and partners too. When the managing partner thanks the bookkeeper who called back to check, the habit sticks.

Ask the same question of whoever runs your IT. If someone calls your provider claiming to be you and asks for a multi-factor reset, what does the provider do to confirm it is really you? A good answer names a specific check. "We'd recognize your voice" no longer counts as one.

Shorter Training, Measured Differently

Kevin Walker of Black Swan described the better approach to TechTarget as short training sprints in place of marathons. For a small office, that can be five minutes at a monthly staff meeting: one real example, one rule, one question. The callback rule fits in a single session.

Training can also fit the job. The front desk needs to hear what a fake caller asking for a patient's records sounds like. Whoever pays the bills needs to practice the vendor bank change call. Anyone with administrator access needs to know that one fake download on their machine can reach every account they manage.

The measurement matters too. A falling click rate on practice phishing emails mostly shows that staff have learned to recognize the practice emails. The more useful numbers are how many people report a suspicious message and how quickly. One fast report lets your IT team pull the same message from every other inbox before anyone else acts on it. Make reporting one click, and make it clear that reporting your own mistake never gets anyone in trouble.

A Rule for AI Tools, and What Regulators Expect

Most offices already have staff using AI tools, approved or not. A one-page policy covers most of the risk. Name the tools staff may use, require that they come from the vendor's official site or app store, and list what never goes into a prompt, starting with patient or client records, passwords and financial data.

The rules many professional offices work under already point this way. The HIPAA Security Rule calls for a security awareness and training program for the whole workforce, management included, and pasting patient information into an AI service that has not signed a business associate agreement may count as an impermissible disclosure. The FTC Safeguards Rule, which applies to tax preparers among many other businesses, requires security awareness training "updated as necessary to reflect risks identified by the risk assessment." A course that still centers on spotting typos may be hard to describe that way. For law firms, the American Bar Association's Formal Opinion 512 asks lawyers to weigh client confidentiality before entering client information into generative AI tools. Cyber insurers may also ask at renewal how staff are trained and how payment changes are verified, and the answers can shape coverage.

Where to Start This Week

If you do one thing this week, write the callback rule and send it to everyone who can move money, change payment details, reset an account or send records. Then ask your IT provider the multi-factor reset question above and write down the answer.

If you are unsure which of your own workflows let a request move money or change access without an independent check, our security assessment traces those approval paths and shows where the rule should apply first. It starts with a 15-minute call and ends with written findings, at no cost and with no obligation.