The Expanding Frontier of Shadow IT: Why Citizen Developers Are the Next Great Security Risk for the Enterprise
Organizations are currently missing a critical target in their governance and training strategies as they navigate the rapid integration of agentic AI into the software development lifecycle. While corporate efforts remain largely anchored in training professional software engineers—a logical reaction to the evolution of the software development lifecycle (SDLC) into the agentic development lifecycle (ADLC)—the most significant and fastest-growing risk surface is emerging from an unexpected corner: the business user. Often referred to as "citizen developers," these non-technical staff members are increasingly utilizing low-code and no-code platforms to build, deploy, and manage applications without the traditional oversight of IT or security departments.
This shift represents a fundamental transformation in how enterprise software is built. As AI tools lower the barrier to entry, employees in departments ranging from marketing and human resources to finance and operations are effectively becoming their own software architects. However, this democratization of development comes with a severe "security literacy" gap. When these users encounter technical hurdles, they often turn to generative AI for guidance. Frequently, these tools suggest deployment environments like Cloudflare or other non-approved platforms, inadvertently creating a massive layer of "Shadow IT." Because these users operate outside the scope of formal security protocols, they often build and ship applications without any training in secure coding practices, threat modeling, or data privacy.
A Chronology of the Low-Code Revolution
The current crisis did not emerge overnight; it is the culmination of a decade-long push for organizational agility. The trend began in the mid-2010s, as companies sought to dismantle the "IT bottleneck" that forced business units to wait months for custom software.
In 2017, a landmark industry report estimated that 60% of custom applications were being developed outside of formal IT departments. At the time, this was viewed as a resounding success, with many enterprises arguing that business departments possessed the domain expertise required to build more effective, user-centric tools than centralized IT teams could. By 2024, that trend had accelerated exponentially. Current projections suggest that 70% of all new business applications and 75% of enterprise-grade software are now built using low-code or no-code tools, with nearly 80% of these users operating entirely outside the reach of IT oversight.
The arrival of Large Language Models (LLMs) and agentic AI in 2022 and 2023 acted as a catalyst for this movement. Where low-code tools previously required some logical understanding of workflow design, AI now allows users to simply describe their needs in natural language, with the AI generating the necessary code, logic, and integration scripts automatically.
The Quantifiable Risk: Security Voids in AI-Generated Code
The productivity gains associated with this shift are undeniable, but the security costs are becoming increasingly apparent. Recent studies from the Cloud Security Alliance (CSA) and various cybersecurity firms have quantified the scale of the danger. According to a 2026 CSA report, AI-assisted code commits are found to expose sensitive information at a rate of 3.2%, compared to just 1.5% for human-written code.
Furthermore, the quality of security controls in these applications is statistically alarming. Independent research has indicated that AI-generated code introduces security vulnerabilities in 45% of development tasks. Perhaps most concerning is the failure rate of basic security controls. When tasked with implementing fundamental protections—such as cross-site request forgery (CSRF) defenses—AI agents have been observed to fail 100% of the time in certain testing environments. Research from Georgetown University’s Center for Security and Emerging Technology (CSET) further underscored this, finding that cross-site scripting (XSS) vulnerabilities were present in 86% of the code samples generated by five major LLMs.
The real-world implications of these failures were laid bare by the security firm Escape, which analyzed over 5,600 publicly available applications built using AI-assisted tools. The audit revealed over 2,000 vulnerabilities, 400 instances of exposed API keys and secrets, and 175 cases of direct Personally Identifiable Information (PII) exposure, including bank account numbers, medical records, and private communication logs.
The Anatomy of an Invisible Breach
The risk is not that these employees are acting with malicious intent; rather, they are performing their jobs as instructed, creating tools that genuinely improve their productivity. The vulnerability arises from the lack of "security awareness" regarding the backend infrastructure of their applications.
Consider a common scenario: a citizen developer in the finance department builds an automated workflow to reconcile invoices. To connect their AI-driven tool to an internal database, the developer generates an API key. The integration works perfectly, and the user moves on to their next task. However, because the developer lacks formal training, there is no visibility into how that key is stored, who can access it, or the potential impact if that key were compromised. If an attacker gains access to that API endpoint, they could potentially move laterally through the enterprise network. Because there is no documentation, log trail, or security oversight, the organization may remain unaware of the breach until a major data exfiltration event occurs.
Strategic Frameworks for Remediation
Industry experts argue that the solution is not to revert to the old model of restricted IT access, but rather to adopt a comprehensive AI literacy framework. Organizations must treat "citizen developer security" as a primary pillar of their cybersecurity strategy.
The AI Adoption Model provides a structured, three-phase approach for managing this risk:
- Phase 1 (AI-Assisted): At this entry level, governance policies must be established alongside basic upskilling. As organizations begin using AI with human supervision, security teams should implement tools to establish performance baselines and monitor the types of data being fed into these models.
- Phase 2 (AI-Native): As teams shift toward more autonomous AI usage, the training must evolve. Users need to be educated not as coders, but as "code reviewers." Even if they are not writing the syntax, they must be trained to identify potential vulnerabilities and verify the security posture of the outputs their agents produce.
- Phase 3 (Agentic Orchestration): In this advanced phase, where multiple AI agents operate in concert, organizations must implement formal policy enforcement, risk scorecards, and audit-ready tracking for every automated commit.
Broadening the Scope of Enterprise Education
The challenge of securing software development is not new. For years, professional developers have struggled with the requirement for secure coding practices, often relying on security teams to catch errors during the final stages of the CI/CD pipeline. However, as the velocity of development increases, the traditional "security gate" model is becoming obsolete.
The imperative now is to move security education out of the IT department and into the broader enterprise. This requires a cultural shift where AI literacy is viewed as a foundational business skill, equivalent to basic data privacy or compliance training. Organizations that fail to extend this education to their non-technical staff are effectively leaving a wide, undefended perimeter.
As agentic AI continues to weave itself into the fabric of daily business operations, the line between "developer" and "business user" will continue to blur. The companies that thrive will be those that empower their workforce to use these tools effectively while providing the guardrails necessary to ensure that increased productivity does not come at the cost of total system compromise. The era of the "amateur developer" is here; the responsibility of the enterprise is to ensure they are no longer operating in the dark.
