Adopting an AI tool adds another system that holds your data and connects to your firm. That is another door, and cyber security is about making sure each door is locked. The discipline is familiar, applied to one more place.
Treat it like any other supplier
Check how the provider secures data, whether it encrypts information in transit and at rest, how it handles access, and what happens in a breach. A provider that cannot answer plainly is telling you something.
Ask the questions a regulated practice has to be able to answer afterwards. Does the contract say client data will not train the provider's models? Where is the data processed and stored? Who inside the provider can reach it, and under what controls? Is there a data processing agreement that meets the UK GDPR? The gap between a consumer tier and a business one is often the gap between a breach and a safe workflow. We set the confidentiality side of that out in client confidentiality when using AI tools.
Ask too who your supplier depends on. Your engagement is with the company whose invoice you pay, and that company buys hosting, screening data, model capacity and search from others your contract never names. Between Wednesday 5 August 2026 and the following Monday, LexisNexis took three Nexis products offline after identifying unusual activity on servers hosted and managed by a third-party vendor. One of them, Nexis Diligence, is the tool compliance teams reach for when they screen a client against adverse media, so a step in onboarding stopped working for the better part of a week. Ask your main suppliers for their subprocessor list in writing, and ask to be told when it changes.
Protect the way in
Use strong, unique credentials and multi-factor authentication on every AI tool, the same as you would for your case management system. Many breaches are not clever attacks but a reused password on a tool nobody was watching.
Limit what it can reach
Give each tool only the access it needs, and review who in the firm can use it. The fewer people and the less data exposed, the smaller the damage if something goes wrong. Keep a record so you can act quickly if a tool is compromised.
That principle is under pressure, because the assistant in most firms has stopped being a reader. On 7 July 2026 Anthropic's Claude Cowork assistant gained write tools on its Microsoft 365 connection, able to send email, manage drafts and calendar events, change mailbox settings and create or update files in OneDrive and SharePoint. An email leaving a solicitor's mailbox is the firm's correspondence whatever drafted it, and a drafting error a person would have caught on the way out of the door now has a route past the door entirely. Treat write access as a decision rather than an upgrade, enable it user by user and prefer draft-and-review over direct sending. The whole question is worked through in when your AI can send email in the firm's name.
Decide how much autonomy you have handed over
The National Cyber Security Centre published interim advice on agentic AI on 20 August 2026, written by one of its principal security architects, and it opens on the observation that several recent incidents have involved AI systems carrying out unsanctioned or unintended activity. Its controls apply in proportion to autonomy, and the greater the autonomy an agent holds, the greater the damage when it malfunctions, reaches information it should not, or acts outside its intended scope.
The safeguards built into a model are a baseline rather than an answer, since they are capable of being bypassed and might not hold in higher-risk settings. And prompting alone is not a defence, so technical and operational limits have to sit behind the instructions. The centre names three patterns of oversight, a human in the loop approving actions before they happen, a human on the loop watching and intervening, and a human out of the loop leaving the agent running alone. Every agent in your firm belongs to one of them.
Before you switch anything on, write down four lines. What the agent is allowed to do without approval, what identity and credentials it holds and for how long, what it can reach across the network, and how you stop it mid task. The fuller version of the question sits in the NCSC advice on agentic AI.
The way in through your front desk
Not every door is technical. The SRA published a substantially revised sectoral risk assessment on 6 August 2026, and its technology section was updated to focus on recent AI risks. It says AI-enabled impersonation techniques, including deepfakes, increase the risk of identity fraud and misrepresentation during client onboarding and throughout the life of a matter.
The voice calling in week nine to change the account for completion money sits inside the same warning as the person who first instructed you. So keep a second channel for anything that moves money, confirm the instruction on a number your firm held before that instruction arrived, and record who confirmed it. Ask your electronic verification supplier, in writing, what its liveness detection does about generated video and about a stream injected between the camera and the software, which are different attacks and are often answered as though they were one. Client due diligence under the Money Laundering Regulations 2017 stays with your firm whether the check runs in your office or inside a supplier's software.
Plan for the day a tool is gone
Security includes availability, and AI earns its place in a firm quietly. Somebody starts using a tool for one task, finds it saves an hour, and spreads the habit without anyone deciding the firm now depends on it. When Fable 5 and Mythos 5 went dark for about nineteen days under export controls, the firms that had routed drafting or research through a single tool found out what that dependence was worth.
Write down, for each tool that touches a regulated step, what the firm does while it is unavailable. Name the fallback, and name the person who decides to switch to it. Keep the manual process alive as well, because a skill you cannot perform is a skill you cannot supervise. Then take the point to your renewals and ask what the supplier owes you when the service stops.
AI tools are safe to adopt when the firm treats them with the same security basics it applies to everything else that holds client data.
The NCSC's small business guide covers the fundamentals in an afternoon.
If nobody has walked your doors since the last tool arrived, a short review locks them while the list is still short: start with a conversation.
