Thanks to AI, the software development industry finds itself at the epicenter of a structural transformation that has irreversibly altered the fundamental notions of integrity, quality and, above all, security and privacy. For decades, systems architecture was based on deterministic rigor in which every line of code was an explicit instruction written by a human with a specific, auditable intent.
However, the emergence and mass adoption of Generative Artificial Intelligence have given way to a paradigm known as "vibe-coding," a methodology where intent prevails over syntax and where probabilistic code generation has replaced technical craftsmanship. This evolution represents a shift in the hierarchy of corporate priorities, instilling a tacit belief in boardrooms and engineering teams: in the race for exponential speed, digital security has gone from being a non-negotiable requirement to a component perceived as optional or secondary.
And the evidence suggests that this phenomenon is creating a technical debt and security crisis of systemic proportions. While productivity appears to scale with the use of autonomous agents, the attack surface expands at a similar rate, driven by context blindness and blind trust in probabilistic models that prioritize immediate functionality over long-term resilience.
That is why recent incidents on high-visibility platforms, such as the Microsoft 365 Copilot flaw and the partial financial collapse of the Moonwell protocol, are not isolated events but symptoms of a deeper pathology in the way the modern digital ecosystem is built and protected.
PETER STEINBERGER, THE AI LUTHIER, JOINS OPENAI TO CLOSE THE CHAT WINDOW
The term vibe-coding, popularized by Andrej Karpathy, describes a state in which the developer "surrenders to the vibes" of AI, completely forgetting the existence of the underlying code to focus on directing intent.
In this ecosystem, natural language has become the new universal programming language. This democratization has allowed individuals without deep technical training to build complete applications in record time, shifting traditional software engineering toward a role of "output curation."
This disintermediation has profound economic implications. The ability of AI agents to replicate user interfaces, dashboards and complex business logic almost instantly has begun to erode the competitive advantage of traditional Software as a Service (SaaS) companies. The phenomenon, dubbed the "SaaSpocalypse," reflects how organizations prefer to build their own internal tools through vibe-coding instead of paying costly subscriptions for user licenses or "seats."
However, this speed has a hidden cost: the elimination of rigorous architectural design processes and the omission of the security testing that usually accompanies manual development.
The erosion of quality and the bias toward functionality over security
A study by Veracode 2025 GenAI Code Security Report offers a worrying view of the quality of AI-produced code. After analyzing more than one hundred large language models across eighty different coding tasks, the research determined that 45% of the generated code contains critical security flaws.
What is most alarming is not only the figure, but the fact that, while AI has dramatically improved its ability to produce code that compiles correctly (reaching success rates of 90%), its ability to guarantee security has remained stagnant.
This divergence between syntactic functionality and security integrity stems from the fact that AIs are essentially prediction engines trained on public repositories containing decades of bad practices, obsolete algorithms and vulnerable patterns. When an AI model faces the choice between a secure but complex method and an insecure but functionally direct one, it chooses the insecure option almost 50% of the time to ensure that the user's application "works" immediately.
AI SOVEREIGNTY ENDS WHERE THE POWER CABLE BEGINS
Vulnerabilities by language and critical patterns
But the security risk introduced by AI is not distributed equally across programming languages. The Veracode report highlights that languages with a long history of public repositories, such as Java, have the highest failure rates due to contamination of training data with old and vulnerable code.
The most common vulnerabilities generated by these tools include logic flaws that are imperceptible to traditional static scanners. For example, AI fails to secure code against Cross-Site Scripting (XSS) in an astonishing 86% of cases and fails to sanitize logs against Log Injection attacks 88% of the time.
These errors stem from "architectural context blindness," that is, the AI knows how to write a function, but it does not understand whether that function will run in an environment that handles sensitive medical data or in a simple personal to-do list.
The Microsoft 365 Copilot flaw
But to better understand this, it is worth looking at some cases of how AI has created complete chaos. In that sense, one of the most significant incidents was the technical vulnerability in Microsoft 365 Copilot, tracked under the identifier CW1226324. This security flaw allowed the AI assistant to bypass the Data Loss Prevention (DLP) policies that corporations rely on to protect their most sensitive information.
First detected on January 21, 2026, the flaw allowed Copilot to read and summarize emails marked with confidentiality labels that should have restricted access to any automated tool. The problem was located in the chat function of the "work tab," which incorrectly processed messages stored in users' Sent Items and Drafts folders.
MICROSOFT PAYS $9.7 BILLION TO IREN FOR ACCESS TO NVIDIA GPUS
This incident sparked a wave of concern in highly regulated sectors. The United Kingdom's National Health Service (NHS) logged the incident, underscoring the real impact on public-sector users who handle confidential citizen data.
Microsoft responded by rolling out a configuration update worldwide in early February, but the trust gap had already opened, and this is just one case among the many that have affected Microsoft recently.
The oracle crisis at Moonwell
Of course, the Microsoft case is an example of what happens in the corporate world, but these events are now also reaching the crypto world. Our case takes us to the Moonwell protocol, which suffered a loss of digital assets valued at $1.78 million due to a configuration error in its price oracle for the cbETH asset. The incident occurred after the execution of governance proposal MIP-X43, which enabled Chainlink OEV wrapper contracts on the Base and Optimism networks.
The technical error was an omission in the price derivation logic: the oracle reported only the gross cbETH/ETH exchange rate instead of multiplying it by the ETH/USD price to obtain the real dollar value. This caused the system to record a price of $1.12 for the cbETH token, when its real market value was approximately $2,200.
What makes the Moonwell case paradigmatic of the current era is the origin of the flaw. Security auditors found that the commits related to the oracle configuration in proposal MIP-X43 came from Anthropic's Claude Opus 4.6 model. The developer responsible used AI to generate the Solidity code, but the lack of integration testing was what allowed an elementary logic vulnerability to reach production.
LESS CLOUD, MORE PRIVACY: FIVE TOOLS FOR RUNNING LOCAL AI IN 2026
This "vibe-coding error" was instantly exploited by liquidation bots. Seeing the price artificially crash to 1.12, the bots attacked the collateral positions, allowing them to seize 1 cbETH by paying an insignificant debt of barely $1. The protocol was left with massive bad debt distributed across various digital assets.