Understanding data security for law firms
David YuI will cover here some practical advice on how you can keep your data secure for your law firm. I think this is particularly important for law firms because they carry some of the most sensitive data I have seen, such as evidence of serious crimes that would cause serious damage to victims and alleged offenders if released publicly, e.g. the Epstein files. I have covered it briefly before here; we can dive a bit deeper. There are two key areas to be mindful of: within your firm, and your software suppliers — basically everyone who has access to your data. It may seem impossible to keep track of where your data moves, but with careful evaluation of your software providers you can hopefully protect yourself and your clients from hitting the news.
Risks from software suppliers
Why you should move to foundation models to reduce risk
If you add a software vendor in front of a foundation model provider you add risk, because it's an additional party involved that has full access to your data. They can advertise any number of certifications and security measures, but you are at their mercy. Generally, smaller companies have less experience and pressure to handle data securely, whereas large companies would have gone through the challenge of keeping data secure from employees and attackers. Additionally, they would have worked with enterprises who demanded secure practices of them.
Many Australian AI legal companies may advertise that their data lives in Australia. However, as the chart above shows, that is not the case, because they need to call downstream dependencies — particularly the foundation models, which are hosted in the US. Some providers, such as Google Gemini, offer multi-region hosting. Enterprise companies may want data to stay in a specific region for compliance.
However, the main security concern is not where the data transits or is stored, but who has access to it. There is a fan-out of people who gain access to your data as you add more providers, since they themselves will use other providers. This doesn't mean you shouldn't adopt new software — it could hugely benefit your business — but you should weigh up the security cost.
The legal AI wrapper problem
This is exactly the shape of the legal AI market today. When I looked at the existing legal AI products, they were expensive, awkward to use, and mostly thin wrappers around the same foundation models you could go to directly. A wrapper is an additional party with full access to your matter files — you pay a markup and widen your data's exposure at the same time.
So I ran an experiment: I encouraged firms to try the Claude app directly instead. The feedback was consistent — better results, lower cost, and one less company holding their data than the legal-specific tool they were already paying for. Part of why AI adoption lags in law is that early legal AI left a poor first impression; the frontier models have improved quickly, and going straight to the source fixes both the experience and the exposure. It's the same principle our own product is built around: the fewer hands your data passes through, the better.
Likelihood of breaches
- Small businesses: 80% suffered at least one cyberattack in 2025, and 59% of SMEs globally reported experiencing a cyberattack in the past 12 months. CNiC Solutions
- Medium and large businesses: in the UK, 67% of medium and 74% of large businesses identified a breach or attack in the last 12 months, versus 43% of businesses overall — detection rises with company size. UK Cyber Security Breaches Survey 2025
- Large enterprises: large organizations recorded 982 security incidents and 751 confirmed breaches in the same period — roughly 4× fewer than SMBs, which logged 3,049 incidents and 2,842 confirmed breaches. Spacelift
The table summarises the 2025 Verizon DBIR breakdown by organisation size, drawn from 22,052 security incidents and 12,195 confirmed breaches worldwide. It shows how heavily smaller organisations are hit.
| Business size | Security incidents (2025) | Confirmed breaches (2025) |
|---|---|---|
| Small (micro/SMB) | 3,049 | 2,842 |
| Large enterprise | 982 | 751 |
Verizon, 2025 Data Breach Investigations Report (DBIR), SMB Snapshot.
Small companies are attacked more often, and the vendor fan-out discussed above only widens their exposure. Each breach is unique, and many go undetected — for example, an employee screenshotting your data. I would estimate the true amounts and severities are larger than what is published.
Mitigations
If you have evaluated the risk and would still like to adopt a new software vendor, you will need to see whether you can mitigate the possibility and impact of a data breach.
Bring your own key (BYOK)
This means you provide your own encryption key to the software provider, and they use it to encrypt your data. It's like putting a password on your data — you need to enter it before you can see it. This requires mature systems to implement, and most small companies can't support it. Additionally, the data is still decrypted when it needs to be used, so you remain exposed to the danger of data being read by employees or logged in plain text in their systems.
Requiring safe data access practices
There are a few practices to ask a provider about
- Just-in-time (JIT) access, where only on-call employees have elevated permissions and any data access must go through a review process.
- An audit trail: if data needs to be accessed for an incident, for example, there should be a clear record of who accessed it and what they did with it.
- Tiered levels or segments of data access — for example, no user has root AWS account access, and the finance department does not have code access.
- Employee computers are monitored for customer data, which is required to be deleted.
- Code changes are always reviewed by another team member. This prevents a single rogue employee from deploying malicious software.
Limited or zero data retention policies
Companies should treat user-generated content (UGC) very carefully, and that means implementing at least a limited data retention policy or, if possible, a zero data retention (ZDR) policy. This means UGC is deleted from all of their systems, e.g. employee computers and servers.
An example of a service that can offer ZDR is an AI provider, because inference is stateless — they only need your data momentarily to process the request and generate the next token. An example of limited data retention is a migration context, where they can delete the data once the migration is done. Some services can't implement either, so you need to treat them carefully — chat applications are one example, since you want the history of messages accessible on any computer.
It's important to check that UGC is not leaked into logs or analytics platforms. It's a common mistake to mishandle user data this way — it's very easy for a developer to add a single line of code that exposes all your data to everyone in their company. They should implement automated scanners to prevent this sort of mistake.
Risk from within your firm
The other side of the coin is your own people this is also very common particularly in non technical professions like law so its important to protect your firm and clients. This risk comes in two forms: malicious and accidental.
Malicious and accidental insiders
A malicious insider deliberately misuses their access — a disgruntled paralegal copying case files before they resign, or someone selling details of a high-profile matter to the press. These cases are rare, but they can be devastating, because employees often have broad, trusted access.
The accidental insider: the employee who emails a document to the wrong recipient, uploads case files to a personal Dropbox to work from home, loses an unencrypted laptop on the train, reuses a weak password, or clicks a phishing link that hands over their credentials. There is no malice involved, but the outcome is the same.
How to reduce the risk
The reason to treat this separately from supplier risk is that the mitigations are different. A supplier you can only evaluate and hold to account; your own staff you control directly, so you can do the following.
- Apply least privilege — give each person access only to the matters they are working on, not the entire document store.
- Require multi-factor authentication (MFA) and full-disk encryption on every device.
- Revoke access promptly when someone leaves; forgotten accounts are a common entry point for attackers.
- Keep an audit log of who opened which client files, so misuse can be detected and traced.
- Train staff to recognise phishing and social engineering, which are the most common ways credentials are stolen.
- Set a clear policy on where client data may and may not go — no personal email, no personal cloud storage, no unapproved apps.
Tools that do the work
You don’t have to build any of this yourself — large firms buy it off the shelf. Endpoint detection tools like CrowdStrike Falcon, SentinelOne, or Microsoft Defender for Endpoint sit on each laptop and flag or block suspicious activity, such as a bulk copy of case files to a USB drive. Identity providers like Okta or Microsoft Entra ID enforce MFA and single sign-on, and let you cut off a departing employee’s access everywhere at once. Device-management tools like Microsoft Intune or Jamf enforce full-disk encryption (BitLocker on Windows, FileVault on Mac) and can remotely wipe a lost laptop. Data-loss-prevention tools such as Microsoft Purview or Google Workspace DLP watch for client data being emailed or uploaded somewhere it shouldn’t go.
This is a lot of work to setup since there are so many tools to evaluate and they also loop back to the software supplier risk discussed above. I would recommend you start with CrowdStrike Falcon and ask employees to setup MFA to your Microsoft accounts. And if your employee loses their laptop just remember to use Microsofts remote wipe feature. Its best to run through a test so you are ready when the scenario arises.
Balancing security and privacy
There is a line to walk here, though: the more closely you monitor staff, the more you intrude on their privacy and their trust. Logging every keystroke, screenshot, and site an employee visits is invasive, hurts morale, and in some jurisdictions is legally restricted. The aim is to monitor the data, not the person — watch for client files moving where they shouldn’t, rather than surveilling everything your staff do — and be transparent about what you monitor and why. Staff who understand the policy are far more likely to follow it than resent it.
Conclusion
The key takeaways are to understand the risk of installing or using software, as it opens a gateway for your data to leave your company. It's best to go with enterprise companies that have a long track record of no data incidents, instead of small vendors who are much more likely to expose your data. You should also understand the software supply chain: small firms end up using large companies' products and infrastructure anyway, so it's best to go straight to the source. For AI products, for example, it's best to use Gemini, ChatGPT, or Claude.
But suppliers are only half the picture. Your own staff are just as likely to be the source of a leak, whether through malice or a careless mistake, so pair your vendor choices with strong internal controls — least-privilege access, MFA, encrypted and monitored devices, and prompt offboarding — enforced with off-the-shelf tools like Okta, Microsoft Intune, and CrowdStrike Falcon. Do this without overreaching into your employees’ privacy: monitor where the data goes, not every move your people make.