The Vibe Coding Hangover: How AI Agents Broke the Database, and How We Fix It

SAN FRANCISCO – November 06, 2026 – We are living through the hangover of the "vibe coding" revolution. Over the past year, the promise of democratized software creation reached its zenith. Anyone with a text prompt could spin up a fully functional web application in minutes using AI builders like Lovable or Bolt.new. It was a triumph of modern commercial strategy and artificial intelligence, removing the friction between an idea and a deployed product.

But as the dust settles on millions of these rapidly prototyped applications, a sobering reality is emerging: the AI agents writing our code are terrible at locking the doors.

Today, Constructive, an infrastructure company backed by Silicon Valley heavyweights like Marc Benioff and Reid Hoffman, released Safegres. It is an open-source, read-only verification tool designed to audit live PostgreSQL databases and catch the exact security misconfigurations that AI agents are currently deploying at an alarming scale. To understand why a highly technical database auditing tool is one of the most critical commercial releases of 2026, we have to look at the structural security crisis currently unfolding beneath the surface of the AI app economy.

The Vibe Coding Security Hangover

The sheer scale of data leakage caused by AI-generated applications is difficult to overstate. Because AI coding agents prioritize rapid functional deployment over rigorous security architecture, they frequently fail to configure Row Level Security (RLS)—the database rules that dictate which specific user can see which specific row of data.

Recent independent investigations have laid bare the consequences. On November 2, the German security firm Crissy Field and the business daily Handelsblatt published a devastating scan of 159,041 projects built with Lovable. They discovered that 24,664 of these sites, utilizing Supabase as their backend, were serving database records to completely anonymous visitors. The exposed data wasn't trivial; it included passport numbers, payment records, contracts, invoices, and over 200,000 private chat messages.

This was not an isolated incident. Cybersecurity firm UpGuard recently found 16,326 Supabase databases exposing readable tables across roughly 300,000 domains, leaking personally identifiable information from US valet services, Canadian immigration firms, and adult creator platforms. Another vulnerability, tracked as CVE-2025-48757, affected over 170 Lovable apps, allowing remote attackers to extract entire tables of user lists and API keys without authentication.

The root cause is a dangerous mismatch between how humans build databases and how machines do. When a human uses a platform dashboard to create a table, security features like RLS are often enabled by default. But when an AI agent writes raw SQL migrations to build the same table, RLS defaults to off. It is a configuration mistake that AI tools are repeating at industrial scale.

The PostgreSQL Battleground for AI Safety

This crisis has forced foundational infrastructure providers into a defensive posture. The database engines built decades ago are now being rapidly retrofitted to survive autonomous agents that spin up thousands of untested schemas before lunch.

The statistics are staggering. According to Databricks' recent "State of AI Agents" report, autonomous agents now generate roughly 80% of all new databases on the Neon platform. Supabase CEO Paul Copplestone recently confirmed a similar metric, noting that approximately 70% of new database activity on their platform originates from AI agents rather than human developers.

Infrastructure providers are scrambling to close the gaps. On October 30, 2026, Supabase took a drastic step, fundamentally changing its default access policies. The platform stopped automatically granting access to new tables in the public schema to anonymous and authenticated roles. Developers—or their AI agents—must now explicitly write GRANT statements to allow access.

However, forcing agents to write explicit permissions only solves half the problem. If an agent writes a permissive grant but still fails to configure RLS, the data remains exposed. The industry desperately needed a way to automatically verify that the access logic written by a machine actually protects the data.

Quality Control for the Agent Factory

This is the exact void Constructive aims to fill with the public release of Safegres. Built on PostgreSQL's own grammar, Safegres is essentially an automated building inspector for database security. It audits RLS policies and grants in a live database, deterministically reporting which roles can reach which rows, without ever touching the application data itself.

"Our tools exist to make whole classes of Postgres vulnerabilities impossible to ship, catching mistakes like missing grants and open RLS policies before they ever become CVEs," said Dan Lynch, Founder and CEO of Constructive. "The same open-source parser behind Safegres already runs in tooling built by teams at Canva and Supabase. With security incidents piling up across AI-built apps, these checks can't be optional. They belong in the development loop, the automated checks that run every time a developer changes code, tested against real Postgres on every commit and migration."

Safegres represents a fascinating shift in commercial software strategy. It relies on Abstract Syntax Tree (AST) parsing—what Constructive calls a "Babel for Postgres." Instead of just looking at the text of the code, Safegres translates SQL into a structured, machine-readable format. It traces inherited roles, views, and shared defaults to figure out exactly what a user can access, executing over 30 security checks and 10 performance checks. Because it relies on deterministic parsing rather than an "AI judge," the tool provides identical, reliable grades every time it runs.

The Autonomous Feedback Loop

Perhaps the most compelling aspect of Safegres is how it fits into the broader evolution of the software development lifecycle. We are moving away from artisanal software crafting and toward software factories, where AI agents act as the assembly line workers.

"With the industry now focusing around software factories, every factory needs quality control at every station," said Anmol Yadav, Head of Infrastructure and Co-founder of Constructive. "Dan and I have been building toward that for years, from replicable deployments and declarative config to treating infrastructure itself as an Abstract Syntax Tree. Safegres puts database security on that same assembly line, so it runs in CI/CD on every commit whether a person or an agent wrote the change."

By integrating directly into Continuous Integration/Continuous Delivery (CI/CD) pipelines, Safegres acts as a hard quality gate. But it also creates a revolutionary feedback loop. When Safegres flags a missing RLS policy or an open grant, it generates a machine-readable report. Coding agents can ingest this report, recognize the security failure, and autonomously write the patch to fix the vulnerability they just created.

As the consumer experience increasingly relies on applications conjured into existence by artificial intelligence, the underlying trust in those applications hinges entirely on data security. The vibe coding era proved that AI can build software at unprecedented speeds. Tools like Safegres are here to prove that we can actually make it safe to use. The era of human-written code may be fading, but the era of machine-audited infrastructure has just begun.