The productivity story is real. Developers using AI coding tools spend less time on mechanical tasks, ship features faster, and — in McKinsey's February 2026 survey of 4,500 developers across 150 enterprises — reduce time on routine coding work by an average of 46 percent. GitHub reports that more than half of all code committed to its platform in early 2026 was generated or substantially assisted by an AI tool. These findings are consistent across multiple independent studies and are not disputed.
The security story is quieter, and less flattering.
ProjectDiscovery's 2026 AI Coding Impact Report found that AI-generated code is outpacing security teams' ability to keep up. Veracode's Spring 2026 GenAI Code Security Update tested over a hundred large language models and found that 45 percent of AI-generated code samples introduce OWASP Top 10 vulnerabilities when no security guidance is explicitly included in the prompt. A 2026 arXiv study analyzing 6,275 public GitHub repositories found that unresolved technical debt climbed from a few hundred issues in early 2025 to more than 110,000 surviving issues by February 2026.
The specific failure modes are consistent across studies. AI-generated code carries 2.74 times more XSS vulnerabilities than human-written code. Eighty-six percent of AI-generated code fails injection defense. Thirty-eight percent of organizations report accidental data exposure via AI-generated code. A Stanford and MIT joint study from March 2026 analyzed over 2 million AI-generated code snippets and found 14.3 percent contained at least one security vulnerability, compared to 9.1 percent for human-written code.
None of this means the tools are broken. It means they were not trained to be security-aware by default, and that the teams deploying them have not always added the layer that makes them so.
The mechanism is not mysterious. An AI coding model optimizes for producing code that compiles, passes the tests available to it, and satisfies the specification it was given. Security constraints that are not part of the specification, or are not enforced by any test in the loop, are not reliably reproduced. The model does not know that the endpoint it is writing should validate its inputs before constructing a SQL query; it knows how to write an endpoint that returns the right response for the test case you provided.
This is the same problem as every other form of code generation, including copy-paste from Stack Overflow: the code that answers the question you asked is not necessarily safe against the inputs you didn't think to test. What changed with AI coding tools is speed and scale. An engineer who spends a Friday generating a month's worth of feature work has also, potentially, generated a month's worth of unreviewed security decisions.
The teams most likely to be sitting on significant unreviewed AI-generated security debt are the ones that adopted fastest — typically startups, lean engineering organizations, and teams with thin security capacity. The teams that moved slowest, who maintained strict code review discipline throughout, have a smaller backlog but also captured less of the productivity benefit. There is no clean position here; the question is whether you know which trade you made.
The research is consistent on what to do: include explicit security requirements in AI prompts, establish human review on security-sensitive paths (input handling, authentication, data access), treat AI-generated code as untrusted until it has passed the same review process as any other contribution, and — if you have not done this — run a targeted audit on the parts of your codebase where AI was most heavily used in 2025 and early 2026.
The tools will improve; the code already shipped will not fix itself.
The short of it.
A cluster of security research published in spring 2026 — ProjectDiscovery, Veracode, Stanford/MIT, and a large-scale arXiv study — consistently finds that AI-generated code introduces security vulnerabilities at a higher rate than human-written code, with 45 percent of AI-generated samples introducing OWASP Top 10 issues when no security guidance is in the prompt. Unresolved technical debt in AI-assisted GitHub repos climbed past 110,000 issues by February 2026. The teams most exposed are the ones that adopted fastest with the thinnest security review process; the fix is explicit security constraints in prompts, human review on sensitive paths, and an audit of what shipped without it.