6 Ways to Prevent Leaking Private Data Through Public AI Tools

Tanya Wetson-Catt • 16 January 2026

We all agree that public AI tools are fantastic for general tasks such as brainstorming ideas and working with non-sensitive customer data. They help us draft quick emails, write marketing copy, and even summarise complex reports in seconds. However, despite the efficiency gains, these digital assistants pose serious risks to businesses handling customer Personally Identifiable Information (PII).


Most public AI tools use the data you provide to train and improve their models. This means every prompt entered into a tool like ChatGPT or Gemini could become part of their training data. A single mistake by an employee could expose client information, internal strategies, or proprietary code and processes. As a business owner or manager, it’s essential to prevent data leakage before it turns into a serious liability.


Financial and Reputational Protection


Integrating AI into your business workflows is essential for staying competitive, but doing it safely is your top priority. The cost of a data leak resulting from careless AI use far outweighs the cost of preventative measures. A single mistake by an employee could expose internal strategies, proprietary code, or sensitive client information. This can lead to devastating financial losses from regulatory fines, loss of competitive advantage, and the long-term damage to your company's reputation.


Consider the real-world example of Samsung in 2023. Multiple employees at the company's semiconductor division, in a rush for efficiency, accidentally leaked confidential data by pasting it into ChatGPT. The leaks included source code for new semiconductors and confidential meeting recordings, which were then retained by the public AI model for training. This wasn't a sophisticated cyberattack, it was human error resulting from a lack of clear policy and technical guardrails. As a result was that Samsung had to implement a company-wide ban on generative AI tools to prevent future breaches.


6 Prevention Strategies


Here are six practical strategies to secure your interactions with AI tools and build a culture of security awareness.


1. Establish a Clear AI Security Policy


When it comes to something this critical, guesswork won’t cut it. Your first line of defence is a formal policy that clearly outlines how public AI tools should be used. This policy must define what counts as confidential information and specify which data should never be entered into a public AI model, such as social security numbers, financial records, merger discussions, or product roadmaps.


Educate your team on this policy during onboarding and reinforce it with quarterly refresher sessions to ensure everyone understands the serious consequences of non-compliance. A clear policy removes ambiguity and establishes firm security standards.


2. Mandate the Use of Dedicated Business Accounts


Free, public AI tools often include hidden data-handling terms because their primary goal is improving the model. Upgrading to business tiers such as ChatGPT Team or Enterprise, Google Workspace, or Microsoft Copilot for Microsoft 365 is essential. These commercial agreements explicitly state that customer data is not used to train models. By contrast, free or Plus versions of ChatGPT use customer data for model training by default, though users can adjust settings to limit this.


The data privacy guarantees provided by commercial AI vendors, which ensure that your business inputs will not be used to train public models, establish a critical technical and legal barrier between your sensitive information and the open internet. With these business-tier agreements, you’re not just purchasing features; you’re securing robust AI privacy and compliance assurances from the vendor.


3. Implement Data Loss Prevention Solutions with AI Prompt Protection


Human error and intentional misuse are unavoidable. An employee might accidentally paste confidential information into a public AI chat or attempt to upload a document containing sensitive client PII. You can prevent this by implementing data loss prevention (DLP) solutions that stop data leakage at the source. Tools like Cloudflare DLP and Microsoft Purview offer advanced browser-level context analysis, scanning prompts and file uploads in real time before they ever reach the AI platform.


These DLP solutions automatically block data flagged as sensitive or confidential. For unclassified data, they use contextual analysis to redact information that matches predefined patterns, like credit card numbers, project code names, or internal file paths. Together, these safeguards create a safety net that detects, logs, and reports errors before they escalate into serious data breaches.


4. Conduct Continuous Employee Training


Even the most airtight AI use policy is useless if all it does is sit in a shared folder. Security is a living practice that evolves as the threats advance, and memos or basic compliance lectures are never enough.


Conduct interactive workshops where employees practice crafting safe and effective prompts using real-world scenarios from their daily tasks. This hands-on training teaches them to de-identify sensitive data before analysis, turning staff into active participants in data security while still leveraging AI for efficiency.


5. Conduct Regular Audits of AI Tool Usage and Logs


Any security program only works if it’s actively monitored. You need clear visibility into how your teams are using public AI tools. Business-grade tiers provide admin dashboards, make it a habit to review these weekly or monthly. Watch for unusual activity, patterns, or alerts that could signal potential policy violations before they become a problem.


Audits are never about assigning blame, but identifying gaps in training or weaknesses in your technology stack. Reviewing logs might help you discover which team or department needs extra guidance or indicate areas to refine and close loopholes.


6. Cultivate a Culture of Security Mindfulness


Even the best policies and technical controls can fail without a culture that supports them. Business leaders must lead by example, promoting secure AI practices and encouraging employees to ask questions without fear of reprimand.


This cultural shift turns security into everyone’s responsibility, creating collective vigilance that outperforms any single tool. Your team becomes your strongest line of defence in protecting your data.


Make AI Safety a Core Business Practice


Integrating AI into your business workflows is no longer optional, it’s essential for staying competitive and boosting efficiency. That makes doing it safely and responsibly your top priority. The six strategies we’ve outlined provide a strong foundation to harness AI’s potential while protecting your most valuable data.



Take the next step toward secure AI adoption, contact us today to formalise your approach and safeguard your business.

Let's Talk Tech

More from our blog

Micro-SaaS Vetting: The 5-Minute Security Check for Browser Add-ons

by Tanya Wetson-Catt • 25 May 2026
Browser add-ons have a funny reputation. They feel “small”. A quick install. A tiny productivity boost. A harmless little helper that lives in your toolbar. But in practice, a browser extension is more like a micro-SaaS vendor sitting inside your browser session. It can see what you see, interact with the pages you open, and sometimes access the same cloud apps your business runs on all day. That’s why a browser extension security check matters. Not because every extension is bad, but because it only takes one over-permissioned add-on or one bad update to turn “helpful” into exposure. The good news is you don’t need a 40-page policy to reduce the risk. A simple five-minute check can prevent most extension problems before they start. Why Browser Extensions Are a High-Leverage Risk Browser extensions sit in the most sensitive place in modern work: the browser tab where your staff live all day. That matters because extensions aren’t just “apps”. They’re granted special authorisations inside the browser. That makes them attractive targets and gives them leverage that’s disproportionate to how “small” they feel. UC Berkeley’s guidance says extensions get “special authorisations,” and the more you install, the bigger the attack surface becomes. The risk is often permission-based. OWASP calls out “permissions overreach” as a core problem. Extensions can request more access than they need, including access to “all tabs, browsing history, and even sensitive user data.” When an extension can read and modify what happens in the browser, it can potentially see data in cloud tools, capture what’s typed into forms, or alter content on a page. It’s also a “change over time” risk. A useful extension today can become a different extension tomorrow. The 5-Minute Browser Extension Security Check This browser extension security check is designed to be fast, repeatable, and realistic. It helps staff make safe decisions in minutes without turning every extension into a big IT ticket. Vet the developer like a real vendor If you wouldn’t give a random supplier access to your customer records, don’t give a random extension access to your browser. Start with the basics: Confirm the developer has a real website, support details, and a consistent name across listings Look for a track record (other products, a clear company presence, updates that look normal) Prefer official stores and trusted sources over “download this .zip” links

LinkedIn "Social Engineering": Protecting Your Staff from Fake Recruitment Scams

by Tanya Wetson-Catt • 22 May 2026
A fake recruiter message is one of the cleanest social engineering tricks around because it doesn’t look like a trick. That’s why LinkedIn recruitment scams work so well inside real businesses. They don’t arrive as malware. They arrive as a normal conversation that nudges someone toward one small action: click this link, open this file, “verify” this detail, move the chat to a different app. A few simple checks, a couple of hard-stop rules, and an easy way to report suspicious outreach can shut these scams down without slowing anyone down. LinkedIn Recruitment Scams LinkedIn recruitment scams artfully blend into normal professional behaviour. The message doesn’t look like a “cyber attack.” It looks like networking, and it borrows credibility from recognisable brands, polished profiles, and familiar hiring language. At platform scale, the volume is also hard to wrap your head around. Rest of World reports that LinkedIn said it “identified and removed 80.6 million fake accounts” at registration from July to December 2024. A LinkedIn spokesperson claimed “over 99%” of the fake accounts they remove are detected proactively before anyone reports them. Even with that level of detection, enough scam activity still leaks through to reach real employees. That’s especially true when scammers tailor their approach to what looks credible in a specific industry and location. The other reason these scams succeed is that they follow a predictable persuasion pattern: urgency, authority, and a quick push to “do the next step.” The FTC describes scammers impersonating well-known companies and then steering targets toward actions that create leverage. These actions include handing over sensitive personal information or sending money for “equipment” or other upfront costs. Once someone is rushed into treating the process as real, the scam doesn’t need to be technically sophisticated. It just needs the victim to keep moving. The Scam Pattern Most Teams Miss 1. A polished approach on LinkedIn The profile looks credible enough, the role sounds plausible, and the message is written in a professional tone. The job post itself may still be oddly generic, though. Amoria Bond notes that fake job postings often “lack details” and lean on broad language to catch as many people as possible. 2. A quick push off-platform The conversation shifts to email, WhatsApp/Telegram, or a “recruitment portal” link. That shift is important because it removes the built-in friction of LinkedIn’s environment and makes it easier to send links, files, and instructions.

3. A credibility wrapper: “assessment”, “interview pack”, or “onboarding”

Airswift flags link/attachment requests and urgency tactics as common red flags. The story is usually something like: “Download this assessment,” “Review these onboarding steps,” or “Log in here to schedule.” Tag Apps Make decisions visible and repeatable by tagging apps. Microsoft explicitly calls tagging apps as sanctioned or unsanctioned an important step, because it lets you filter, track progress, and drive consistent action over time. 4. The pivot: money, sensitive info, or account takeover Scammers impersonate well-known companies and then ask for things legitimate employers typically don’t: payment for “equipment” or early requests for personal information. Another variation is more subtle: “verification” steps that are really designed to steal identity details or compromise accounts. 5. Pressure to keep moving If someone hesitates, the scam leans on urgency: “limited slots,” “fast-track hiring,” “complete this today.” That’s why Forbes frames the key skill as slowing down and checking details, because the scam depends on momentum. Red Flags Checklist for Staff Here are the red flags to look out for. Red flags in the job posting The role is oddly vague or overly broad. Generic responsibilities, unclear reporting lines, and “we’ll share details later” language are common in fake listings. The company's presence doesn’t match the brand name. Thin company pages, inconsistent logos/branding, or a web presence that feels incomplete are worth pausing on. The process is “too easy, too fast.” If the listing implies immediate hiring with minimal steps, treat it as suspicious. Red flags in recruiter behaviour They push you off LinkedIn quickly. Moving to WhatsApp/Telegram or personal email early is a common tactic. They use a personal email address or unusual contact details. Be specifically cautious of recruiters using free webmail accounts instead of a company domain. They avoid verification. If they dodge basic questions, treat that as a signal, not a scheduling issue Hard-stop requests Any request for money or fees. Application fees, equipment purchases, “training costs”, gift cards, crypto, that’s a hard stop. Requests for sensitive personal info early. Bank details, identity documents, tax forms, or “background checks” before a real interview process is established. Requests for verification codes. If anyone asks you to read back a one-time code sent to your phone/email, assume they’re trying to take over an account. Requests for non-public company information like org charts, internal system details, client lists, invoice processes and security tools. Look out for requisitions for anything beyond what a recruiter would reasonably need. Stop Scams With Simple Defaults LinkedIn recruitment scams don’t succeed because staff are careless. They succeed because the outreach looks normal, the process feels familiar, and the next step is always framed as urgent. The fix isn’t turning everyone into an investigator. It’s setting simple defaults that make scams harder to complete: slow down before clicking, verify the recruiter and role through official channels, keep conversations on-platform until identity checks out, and treat money requests, code requests, and early personal data demands as hard stops.  When those habits are standardised, the scam loses its leverage.

The "Legacy Debt" Audit: Identifying the 3 Oldest Risks in Your Server Room

by Tanya Wetson-Catt • 18 May 2026
The most dangerous thing in a server room is often the phrase, “Don’t touch that.” It’s usually said with a half-joke and a grimace. It refers to the old box that “still works”, runs something important, and has survived so many fixes and workarounds that nobody feels confident changing it anymore. That’s legacy debt. Not just “old tech”, but old tech that’s become a dependency. It’s the kind that quietly accumulates risk until it turns into downtime, security exposure, or an emergency upgrade at the worst possible time. A legacy debt audit is the fast way to bring that risk back into the light. What Legacy Debt Really Looks Like Legacy debt isn’t “old gear”. It’s old gear that has become normal. It’s the server that runs a critical app, the edge device nobody remembers buying, the workaround that turned into a dependency. Over time, that debt stacks up quietly. Infinite Lambda describes legacy debt as something that “happens even to the best systems,” “silently accruing costs and constraints,” and it can “accumulate basically unnoticed until it is too costly to ignore.” That’s why a legacy debt audit isn’t a theoretical exercise. It’s a visibility exercise to bring the oldest, highest-leverage risks back onto the list of things you actively manage. The security problem shows up when “old” becomes “unpatchable.” The UK’s NCSC guidance on obsolete products says, “Ideally, once out of date, technology should not be used,” and “the only fully effective way to mitigate this risk is to stop using the obsolete product.” If something can’t be updated, weaknesses don’t age out. They sit there, waiting for the wrong day. Legacy debt also looks like basic server hygiene slipping. NIST SP 800-123 frames secure server operations as an ongoing process: “Maintaining the secure configuration through application of appropriate patches and upgrades, security testing, monitoring of logs, and backups…” It also calls out foundational hardening steps like “Patch and upgrade the operating system” and “Remove or disable unnecessary services, applications, and network protocols.” When those basics become inconsistent, legacy debt turns into a reliability and incident-response problem, not just a security one. Finally, legacy debt often hides at the edge. If you have end-of-support internet-facing devices, you’ve got high-leverage risk in the most exposed place. The 3 Oldest Risks to Find First These three categories are where “old” most often turns into outsized risk, because they combine age with leverage: they either sit at the front door, can’t be fixed anymore, or have quietly drifted out of a safe baseline. Risk #1: End-of-support edge devices If you’re looking for high-leverage legacy debt, start at the edge. Firewalls, VPN gateways, routers, and other internet-facing devices are the front door to your environment. When they reach end-of-support (EOS), they don’t just become outdated. They become harder to defend because security fixes stop arriving. What to check in your audit List every edge device (firewall, VPN, router) and the support status for each one Confirm which ones are internet-facing and which services are exposed Identify devices that can’t run the current firmware or no longer receive updates. Risk #2: Obsolete products that can’t be fixed anymore Obsolete products are the purest form of legacy debt: things that are still operating but no longer receive security updates. That means every new vulnerability becomes permanent. In other words, there’s no clever workaround that makes an unsupported system “safe”. There are only risk reductions until you can replace it. What to check in your audit Identify anything past support: server OS versions, appliances, old hypervisors, and line-of-business apps Flag systems that require exceptions, like the ones with old protocols, weak auth, and special firewall rules Find the “business-critical but unsupported” systems. Risk #3: “It still works” servers with neglected basics This is the sneakiest risk because it looks normal. The server is supported. The hardware runs. Nobody’s complaining. But the basics have drifted: patching is inconsistent, unnecessary services are still running, and backups haven’t been proven under pressure. SP 800-123 Guide to General Server Security frames secure server operations as an ongoing discipline, including “patches and upgrades,” “monitoring of logs,” and “backups.” It also calls out core hardening steps like “Patch and upgrade the operating system” and “Remove or disable unnecessary services, applications, and network protocols.” Those are the unglamorous fundamentals that stop small problems from turning into long outages. What to check in your audit Patch reality: what’s the current patch level and how often do updates slip? Service sprawl: what’s running that doesn’t need to be running? Admin and service accounts: where are the broad permissions and shared credentials? Backup confidence: when was the last restore test and did it succeed? Change control: who can make changes, and how are they tracked? Stop Carrying Silent Risk Legacy debt doesn’t announce itself. It sits quietly in the background until the day it becomes downtime, exposure, or an emergency upgrade you didn’t plan for. A legacy debt audit gives you control back by turning “we should deal with that someday” into a shortlist you can act on. Start with the highest-leverage risks: end-of-support edge devices, obsolete products that can’t be patched, and servers where the basics have drifted. Then assign owners, set dates, and move one item at a time from “too scary to touch” to “handled”.  Contact us for help running your next legacy debt audit.