Is Vibe-Coded Software Safe to Use in Production? (2026)
If your internal tool is now used by paying customers, the code did not change overnight, but the people it has to protect did. It was built, by you, a contractor or an AI tool, for staff who were trusted to see everything. Now outsiders log in, pay you, and expect their data to be theirs alone. This guide covers what changes: keeping each customer’s data separate, sign-in, an audit trail, backups, monitoring, support, contracts, and the order to fix them in. I am Paul Byrne, President of Razoyo, a software engineering firm that has taken two internal hotel tools into platforms other companies use, so weigh the part about us accordingly.
Start by proving that one customer cannot see another customer’s data, because that is the failure an internal tool is most likely to have and the one most likely to cost you the customer. Then add stronger sign-in, get keys out of the code, restore a backup once, and set up a monitor so you hear about an outage before a customer does. The audit trail, support path and contract paperwork come next.
The worry is justified. An internal tool is built on trust: everyone who logs in works for you, so nobody asked what one user should not see. Our Matrix case study describes the internal hotel sales CRM we rebuilt for M1 Intel in those terms: “The original system was built for a single organization. Every piece of state, every data model, every workflow was scoped to one tenant.” That is normal until a second company logs in. OWASP’s Multi-Tenant Application Security Cheat Sheet puts the stakes in one line: “a single vulnerability can expose all tenants’ data.”
If a customer has already reported seeing someone else’s records, treat it as an incident now: switch off the feature involved, work out what was exposed, and get legal advice on who you have to tell.
Who can reach the data changes, even though the screens look the same. Here is each change, its risk and the first fix.
| What changed | The risk | The first fix |
|---|---|---|
| Several companies share one app | One customer sees another's records | Try to reach one test customer's data as another |
| Strangers hold logins | A stolen or shared password opens an account | Multi-factor sign-in, admin accounts first |
| Keys and passwords sit in the code | Anyone who has seen the code can reach your accounts | Move them to hosting settings and replace them |
| Customers' data is now your responsibility | A bad change or an attack deletes their data | Back up off the hosting, then restore once |
| Downtime costs someone else money | Customers find outages before you do | An outside check that alerts a named person |
| Disputes need evidence | "Who changed this?" has no answer | Log who did what, when, for which customer |
| Every release reaches paying users | A Friday fix breaks a customer's Monday | A test copy and a quick way back |
| Customers expect answers | Everything lands on the person who built it | One support inbox with a named owner |
| Contracts arrive | You sign promises the tool cannot keep | Check every data clause against the tool |
Make the server decide which customer every request belongs to, from the login, and check it on every request.
OWASP’s Authorization Cheat Sheet says permission “should be validated correctly on every request,” and that an application “should be configured to deny access by default.” One miss is enough: “Remember an attacker only needs to find one way in.” An internal tool’s default was the opposite. If you could log in, you could see it.
Screens are rarely where the leak is. Check these too:
Roles matter as much as the boundary. Our Spider case study describes how the platform now part of Lighthouse serves many management companies at once: “Users are assigned roles on specific hotels or portfolios, so each management company’s team sees only its own properties while every company runs on one shared platform.” Give each customer an admin who adds and removes their own people. For a pass or fail version, run the Second-Customer Test in our roundup of firms that build multi-tenant SaaS platforms.
Every outside person needs their own login, multi-factor authentication, and a way to be switched off the same day they leave. CISA’s guidance on multifactor authentication is direct: “Users who enable MFA are significantly less likely to get hacked.” Start with your own admin accounts, which reach every customer.
Then check the flows around it:
Keys belong in the same step. Our production-readiness scorecard on the prototype to production page defines security and access as “An auth model, least-privilege access, and secrets kept out of the code.” Move each password and key into the hosting platform’s settings and replace it, one at a time, on a test copy first.
Yes, because the first dispute will ask who changed what, and memory is not evidence. OWASP’s Logging Cheat Sheet says logs must record “when, where, who and what” for each event, and lists audit trails such as “data addition, modification and deletion, data exports.” The multi-tenant cheat sheet adds the customer: “Include server-verified tenant context in tenant-scoped security and audit events.”
Start with sign-ins, failed sign-ins, permission changes, exports, deletions and staff access to customer accounts. Keep passwords and payment details out of the logs. Our manual operations page names the cost of skipping it: “The second-order costs are the real ones: error rate, no audit trail, and a process that stops the day someone’s out.” Spider “keeps a daily history of every forecast and rate recommendation it produces,” so when a customer asks why a number was what it was, the answer is in the system.
Back up outside the app’s own hosting, then prove you can restore. CISA’s #StopRansomware Guide asks organizations to “regularly test the availability and integrity of backups in a disaster recovery scenario.” A backup you have never restored is a hope.
Multi-tenancy adds a wrinkle: “A shared backup can contain data from several tenants,” as the OWASP cheat sheet notes. Restoring the whole database to fix one customer’s mistake rolls back everyone else. Know how you would recover one customer alone. Then write down how much data you could lose and how long a restore takes, because contracts will ask. Our scorecard’s standard is “Data that stays correct under concurrent use, validated, backed up, recoverable.”
Set up an outside monitor that checks the login and the main task every few minutes and alerts a named person. Our scorecard’s line is “You find out it’s down from a monitor, not from a customer.”
Then watch speed, not only uptime. Google’s Site Reliability Engineering book says “The four golden signals of monitoring are latency, traffic, errors, and saturation.” The Matrix case study describes the old system: “The existing platform struggled under load, with slow queries and UI lag making the product frustrating to use.” If your tool is already falling over, read what to do when a launched app keeps going down.
Decide before the first ticket: one support inbox, a named owner, the hours you cover, and what counts as urgent. Otherwise every question lands on the person who built the tool.
This is general information, not legal advice: once customers pay, your terms, their contracts and the law where they operate can all put duties on how you handle their data, so have a lawyer read them against what the tool actually does. The risk is signing a promise the code cannot keep.
What usually shows up:
Health, payment and children’s data carry their own rules. If your tool holds any, ask a lawyer before the next contract.
Do these in order. Each step comes from our seven-dimension production-readiness scorecard or from what the Matrix and Spider rebuilds needed. Start steps 1 to 3 this week.
Architecture is missing on purpose. Whether the structure can carry outside customers is a judgment that needs someone who can read the code.
It turns on one question: does the data model assume a single organization all the way down? If customer separation can be added to a sound structure, harden it. If every table, query and workflow was written for one company, the foundation has to change.
Our harden or rebuild page sets out the signs. A hardening case has “Missing controls that can be added”; a rebuild case has “Trust boundaries in the wrong place” and “A ceiling that gets worse as volume grows.” And: “Most systems are a mix. That is why the decision is made on an audit, not a guess.”
Matrix needed a rebuild: the team “made the harder call: tear it down and build something that could actually grow.” Spider was a system that hit its ceiling. One of its four limits was access: “Revenue managers needed to use it directly, not through a consultant.” It went from an internal tool that emailed reports to a self-service platform used by more than 10,000 people a day, and we are still its engineering team. Spider, the hotel business intelligence platform now part of Lighthouse, ranked #1 of 74 for Business Intelligence at the 2025 HotelTechAwards.
If your tool started as a spreadsheet, see replacing a business-critical spreadsheet and our spreadsheets to platform page.
We start by looking, not by quoting a rebuild. The first step is a free 30 minute consultation with a senior person. If it is worth going further, a System Spec is $2,500 fixed and time-boxed: it checks what you have against seven production-readiness dimensions and gives you a written harden-or-rebuild call with a senior-engineer estimate. The document is yours whether or not you hire us.
Production builds start from $10,000, ongoing support from $1,500 a month, and dev bursts are $1,200 a day or $4,500 a week. Every price is on our pricing page and on our process page. Above the Spec, a business analyst puts a level of effort and a budget on each piece of work and gets your confirmation first. No approval, no work. If this tool is one of several systems you mean to build properly, see our custom systems page.
More often than a firm like ours would like:
Our internal tool is now used by paying customers and we are worried about security. What should we do first? Prove that one customer cannot see another customer’s data: log in as one test customer and try to reach a second one’s records through screens, links, exports and search. Then add multi-factor sign-in, move secrets out of the code, restore a backup once and set up an outside monitor.
How do I check whether one customer can see another customer’s data? Create two test customers. Log in as the first and try to open the second one’s records by changing the ID in the address bar, running exports, searching and opening uploaded files. If anything of the second customer’s appears, separation is not enforced on the server.
Do we need to rebuild our internal tool to serve outside customers? Not always. If customer separation can be added to a sound structure, harden it. If every table, query and workflow assumes one organization, rebuilding the core is often the honest answer. That is why Razoyo rebuilt an internal hotel sales CRM into Matrix, a multi-tenant SaaS platform, for M1 Intel.
Do customer accounts need multi-factor authentication? Start with every admin account, especially any that can see all customers, then offer it to every user. CISA says users who enable MFA are significantly less likely to get hacked.
What should an audit trail record in a multi-tenant tool? When, where, who and what for each important event, plus which customer it touched: sign-ins, failed sign-ins, permission changes, exports, deletions and staff access to customer accounts. Keep passwords and payment details out of the logs.
How should we handle backups once customers depend on the tool? Back up outside the app’s own hosting, restore into an empty database to prove it works, and know how you would recover one customer’s records without rolling back everyone else.
What will customers ask about security and data in their contracts? Usually a security questionnaire, where their data is stored and who can see it, what happens to it when they leave, how fast you will report an incident, and what uptime you commit to. This is general information, not legal advice, so have a lawyer check each clause against what the tool really does.
How much does it cost to get an internal tool ready for paying customers? At Razoyo the first conversation is free and a System Spec is $2,500 fixed, with a written harden-or-rebuild call and an estimate you keep. Production builds start from $10,000 and support from $1,500 a month.
If customers already log into a tool you built for your own team, that is a free 30 minute conversation. Bring your two-customer test result, even half done. We will tell you which steps matter most, whether anything is urgent, and whether a System Spec is worth it. Schedule Free Consultation.
Paul Byrne is President of Razoyo, a software engineering firm in The Colony, Texas taking prototypes and constrained systems into production since 2011.
These links open AI platforms with pre-written prompts about this page.
We use cookies to improve your experience. Do you accept?
To find out more about the types of cookies, as well as who sends them on our website, please visit our cookie policy and privacy policy.