Perspective
On this page — jump to the part you need:
Artificial intelligence creates security risks that traditional cybersecurity controls were not designed to address.
Organizations adopting AI must consider not only whether their systems are secure, but also what their AI systems can access, what information employees can provide to them, how AI-generated outputs are used, which actions AI agents are permitted to take, and how those activities are monitored and governed.
AI governance and cybersecurity are becoming inseparable.
Vanbreed Consulting Group helps organizations address both sides of the problem: establishing the technical safeguards that protect systems and data while building the governance controls that define how artificial intelligence can be used safely, responsibly, and effectively.
The goal isn't to prevent organizations from using AI. It is to create the technical and governance framework that allows them to use it with confidence.
Traditional cybersecurity programs are built around relatively predictable systems. Applications have defined interfaces. Users have permissions. Databases have access controls. Security teams can identify where sensitive information resides and establish rules governing who can access it.
AI complicates that model.
A large language model can receive unstructured information, combine information from multiple sources, generate new content, call external tools, retrieve enterprise data, and increasingly take actions through AI agents.
That means the security boundary is no longer limited to the application.
Organizations must consider the model, prompts, data, retrieval systems, integrations, tools, agents, users, outputs, and downstream actions as part of the security architecture.
Governance determines what should be allowed.
Cybersecurity helps enforce those decisions.
Organizations sometimes approach AI security as a question of whether an AI model itself is secure. In practice, many of the most significant risks exist around the model.
Employees can unintentionally provide confidential, proprietary, regulated, or personally identifiable information to AI systems.
Organizations need controls governing which AI platforms may be used, what information may be submitted, where that information is processed, and how providers retain or use it.
Employees don't necessarily wait for an enterprise AI strategy.
Public AI tools can quickly become unofficial productivity platforms for writing documents, analyzing information, generating code, summarizing customer data, or solving technical problems.
Organizations therefore need to understand which AI systems are being used and establish practical policies that distinguish acceptable experimentation from unacceptable risk.
Simply banning AI often moves the activity out of sight rather than eliminating it.
An AI application connected to enterprise data may technically have access to far more information than an individual user should be able to retrieve.
A secure AI architecture must ensure that existing authorization boundaries continue to apply when AI becomes another way of accessing information.
The question isn't simply:
“Can the AI access this information?”
It is:
“Should this AI be allowed to access this information on behalf of this particular user for this particular purpose?”
AI systems can receive instructions from more places than the person using them.
Documents, websites, retrieved data, emails, and other external content can contain instructions intended to manipulate an AI system's behavior.
When AI is connected to enterprise systems or tools, prompt injection becomes more than an output-quality problem. It can become a security problem.
The risk changes significantly when AI moves from generating information to taking action.
An AI assistant might draft an email.
An AI agent might send it.
Another agent might update a customer record, execute a workflow, create an account, modify a document, interact with an API, or initiate a business process.
The more authority an AI system receives, the more important traditional security concepts such as least privilege, separation of duties, authorization, monitoring, and approval controls become.
AI governance establishes the boundaries of that authority.
Cybersecurity helps make those boundaries enforceable.
Organizations are rapidly acquiring software products that contain embedded AI capabilities.
Those capabilities introduce questions that traditional vendor assessments may not fully address:
AI therefore becomes part of both cybersecurity risk management and vendor governance.
One of the biggest gaps in AI governance occurs between policy and implementation.
A policy might say:
“Employees may not submit confidential customer information to unauthorized AI systems.”
That's a governance rule.
But several technical questions immediately follow.
How does the organization identify authorized AI systems? How are users authenticated? What data can those systems access? Can sensitive information be detected or blocked? Are prompts and responses logged? Who reviews violations? What happens when an AI application accesses information through an API rather than through a user interface?
Governance without implementation can become a document nobody can enforce.
Security without governance can produce technical controls without clear business rules.
Effective AI risk management connects the two.
VCG approaches AI security as an architectural problem rather than treating the model as an isolated component.
An enterprise AI solution may include:
Users → Applications → Identity → AI Models → Retrieval Systems → Enterprise Data → APIs → Tools → Agents → Business Systems
Every connection introduces a trust decision.
Security architecture should address those decisions deliberately.
Depending on the environment, controls may include:
The appropriate controls depend on what the AI system actually does and the consequences of failure.
AI governance should not assume that every AI decision requires human approval.
That would eliminate much of the value of automation.
Instead, organizations should determine where human oversight is necessary based on risk.
A low-risk AI system generating an internal meeting summary may require little oversight.
An AI system recommending whether a customer receives a service, changing financial information, accessing sensitive records, or initiating an external action may require considerably more.
The important question is not whether a human is involved.
It is whether the level of oversight matches the potential impact of the AI system.
Organizations also need to consider what constitutes an AI-related security incident.
Traditional incident response focuses on events such as unauthorized access, malware, credential compromise, or data exfiltration.
AI introduces additional scenarios.
An AI system may expose information to an unauthorized user. An agent may take an inappropriate action. A retrieval system may return information outside the user's permissions. A third-party AI provider may change how customer information is processed. A prompt-injection attack may manipulate an AI-enabled workflow.
AI incident response should therefore connect existing cybersecurity procedures with AI governance.
Organizations need to know:
What happened?
What data was involved?
What model or system was involved?
What actions did the AI take?
Who or what authorized those actions?
Can the activity be reconstructed?
What controls failed?
How do we prevent recurrence?
That requires logging and auditability to be designed into AI systems before an incident occurs.
Organizations do not need to invent an entirely new risk-management system for artificial intelligence.
AI governance can extend existing cybersecurity, privacy, risk, and compliance programs.
Depending on the organization and regulatory environment, that may include frameworks and guidance such as:
NIST Cybersecurity Framework (CSF)
Provides a structure for managing organizational cybersecurity risk.
NIST AI Risk Management Framework (AI RMF)
Provides a framework for identifying, measuring, managing, and governing risks associated with artificial intelligence.
NIST Generative AI Profile
Extends the AI RMF with considerations specific to generative AI.
NIST Secure Software Development Framework (SSDF)
Provides practices for integrating security throughout the software development lifecycle.
OWASP Top 10 for Large Language Model Applications
Identifies common security risks affecting applications built around large language models.
ISO/IEC 27001
Provides an established framework for information security management systems.
ISO/IEC 42001
Provides a management-system framework specifically for artificial intelligence.
Rather than treating these as competing standards, organizations can map AI-specific controls into the governance and cybersecurity programs they already operate.
Vanbreed Consulting Group combines AI architecture, software engineering, cybersecurity, data engineering, and AI governance to help organizations move from AI experimentation to controlled enterprise adoption.
Our work can include:
Organizations face two bad extremes.
One is adopting AI rapidly without understanding the security, privacy, operational, and governance consequences.
The other is responding to uncertainty by preventing employees and business units from using AI at all.
Neither creates a sustainable AI strategy.
Effective AI governance establishes where AI can be used, what it can access, what it can do, how its activity is monitored, and when human oversight is required.
Cybersecurity provides many of the technical mechanisms necessary to enforce those decisions.
Together, they allow organizations to move beyond uncontrolled experimentation toward secure, governed, production-ready AI.
Whether your organization is establishing its first AI governance program, evaluating existing AI usage, securing an enterprise AI application, or preparing to deploy AI agents, Vanbreed Consulting Group can help identify the risks and build practical controls around them.
Start with understanding what your AI can access, what it can do, and what happens when it gets something wrong.