An AI-built application may work exactly as promised and still expose business data, use excessive permissions or fail in ways its creator never tested.
Asking AI to build something securely is a requirement. It is not evidence that the finished application is secure.
Before connecting a new application to company systems, a business should ask who reviewed it, what was tested, what it can access, how its activity is logged and who remains responsible after it launches.
The lights came on. But is the wiring safe?
Imagine a very good plumber decides to rewire your office.
They ask AI how to do it, follow the instructions, flip the breaker and the lights come on.
Maybe they did everything correctly. But before connecting the entire building, most of us would still want an electrician to inspect the work.
The lights turning on proves the wiring works. The inspection helps prove it is safe.
That is the concern with some of the AI applications suddenly appearing in businesses.
The person who created the application may understand the business problem extremely well. They may also have built something genuinely useful.
But prompting an application into existence does not automatically give someone the experience needed to evaluate its architecture, code, permissions, dependencies or potential failure modes.
What is vibe coding?
Vibe coding generally describes building software by explaining the desired result to an AI coding tool and then repeatedly refining what it produces.
The term should not be used to dismiss every developer who uses AI. Experienced developers use AI to write code, generate tests, find bugs and reduce repetitive work.
The risk appears when the creator cannot independently explain or review what the AI produced but still puts the application into production or connects it to live business data.
That creates an important question:
Who is qualified to recognize what the AI may have missed?
A working application is not a security review
A successful demonstration usually follows the happy path.
The right person signs in. The expected data is available. Every connected system responds normally. The application does what it was asked to do.
Security failures often live outside that path.
- Can one user see information belonging to another user or customer?
- Can the application take a sensitive action without human approval?
- Are passwords, API keys and access tokens stored securely?
- Does the application have broader permissions than it needs?
- What happens when a connected service, AI model or API changes?
- Are security-related actions logged well enough to investigate a problem?
Telling an AI tool to use secure development practices is a reasonable starting instruction.
It is not independent validation that the finished application meets those requirements.
What the recent outages actually tell us
The recent Microsoft 365 and FAA disruptions naturally led people to ask whether AI was involved.
There is no public evidence that AI caused either incident.
Microsoft's preliminary findings described a software defect that became serious when traffic and available capacity changed. The FAA said legacy equipment became unstable when excessive connection attempts flooded its network during work that included a new router deployment.
The useful connection is not that AI caused the outages.
It is that complicated systems can behave very differently under real-world conditions than they did during a test.
Changes in traffic, hardware, dependencies and connected systems can expose a failure nobody saw during the demonstration.
If organizations with experienced engineering teams and formal testing processes can miss an unusual failure mode, why would we expect a quickly built AI application to require less review?
Related Tech Rage IT coverage:
AI is also helping researchers find serious vulnerabilities
AI is not only being used to create software. It is also helping security researchers find vulnerabilities faster.
In October 2026, the PHP project released fixes for two critical vulnerabilities that could allow unauthenticated remote code execution under the right conditions.
OpenAI and the PHP Foundation described AI-assisted work as part of the research and validation process. Human maintainers and security professionals still had to determine which reports were valid, develop the patches and coordinate the release.
That human involvement matters.
The PHP Foundation reported that its security team had received more than 100 reports that ultimately proved to be invalid, duplicates or not reproducible. AI helped with the research, but people made the decisions and completed the remediation.
In November, a researcher publicly released more than 30 alleged vulnerabilities affecting N-able's N-central platform. N-able responded with multiple critical hotfixes strengthening authentication, permission checks, Agent and Probe communications, file handling and other security controls.
These examples do not mean PHP or N-central was vibe coded. They also do not prove that AI-generated software is automatically insecure.
They demonstrate that serious vulnerabilities can exist in complicated software—and that finding a problem is only one part of the work.
Validation, remediation, maintenance and accountability still require a real process.
Read more: How AI helped researchers uncover critical PHP vulnerabilities
Eight questions to ask before trusting AI-built software
1. Who reviewed the design and code?
The reviewer should have enough development and security experience to understand what the application is doing. Asking the same AI that created the code whether its work is secure is not an independent review.
2. What security testing was performed?
The answer should match the application's potential risk. Testing may include code analysis, dependency scanning, authentication and authorization testing, abuse-case testing and independent penetration testing.
3. What can the application access?
Document every connected system, account and data source. The application should receive only the permissions it genuinely needs.
4. Which actions require human approval?
Sending an email and deleting customer records are not the same level of risk. Define where a person must review or approve an action before it occurs.
5. How is business data handled?
Know where the data goes, how long it is retained, whether it may be used to train a model and how it can be deleted.
6. What is logged and monitored?
The business should be able to determine what the application did, which identity it used and whether an action failed or behaved unexpectedly.
7. Who maintains it?
AI models, software libraries, APIs and integrations change. Someone must be responsible for monitoring vulnerabilities, testing updates and keeping the application working safely.
8. What happens when something goes wrong?
Identify who owns containment, investigation, customer notification, recovery and any required reporting before an incident occurs.
What practical AI governance looks like
AI governance does not need to mean a complicated approval committee or a policy nobody reads.
For most small and midsized businesses, it means:
- Knowing which AI tools are being used
- Deciding which tools are approved
- Reviewing vendors before granting access
- Limiting permissions
- Establishing human approval boundaries
- Training employees
- Reviewing tools as their capabilities change
It also means resisting the urge to connect every AI tool to everything the business uses simply because the integration exists.
A capability is not automatically a good business decision.
The process should be lightweight enough that employees will follow it and strong enough that "it worked in the demo" is not the final security review.
What should businesses do next?
Start by listing the AI tools and AI-built applications already being used inside the business.
For each tool, document:
- What it can access
- Who approved it
- What business data it handles
- Which actions it can take
- Whether its activity is logged
- Who is responsible for maintaining it
Then use the eight questions above to decide which tools need a deeper review.
At Tech Rage IT, we believe practical AI governance is the next step in responsible AI adoption. The goal is not to stop useful experimentation. It is to make sure businesses understand what—and who—they are trusting.
Frequently asked questions
Is vibe coding safe?
It can be, but the development method does not prove that the result is secure. The risk depends on what the application does, what it can access and whether qualified people reviewed and tested it before use.
Can AI test its own code?
AI can generate tests, review code and identify potential vulnerabilities. Those capabilities are useful, but they should not be the only validation for an application that handles sensitive data or performs important business actions.
What is AI governance for a small business?
AI governance is a practical process for knowing which AI tools are in use, deciding which tools are approved, limiting access, establishing approval boundaries, training employees and reviewing those tools over time.
Should businesses avoid AI-built applications?
No. Businesses should evaluate them based on their access and potential impact. A low-risk internal tool may need a lighter review than an application connected to customer records, financial information or administrative accounts.
What proof should an AI software vendor provide?
At minimum, the vendor should be able to explain how the application was reviewed and tested, what it can access, how it handles data, how activity is logged, how vulnerabilities are managed and who is accountable after launch.
About Matt Rose
Matt Rose is the CEO and co-founder of Tech Rage IT, a Central Florida managed IT and cybersecurity company. He regularly helps businesses understand technology risk, cybersecurity and practical AI adoption.
Last reviewed: November 19, 2026