Orca Security Team · 10 minute read

Table of Contents

  • Key Takeaways 
  • What Is Vibe Coding and Why It Creates a New Security Problem
  • The Biggest Vibe Coding Security Risks for Engineering and Security Teams
  • Common Vibe Coding Security Vulnerabilities in AI-Written Code
  • Real-World Vibe Coding Security Issues: What Has Already Gone Wrong
  • How to Secure Vibe-Coded Apps From Code to Cloud
  • Best Vibe Coding Security Tools in 2026 (and What to Look For)
  • Building Guardrails: A Vibe Coding Security Policy for Your Organization
  • How Orca Connects Vibe Coding Security to Live Cloud Exposure
  • Frequently Asked Questions about Vibe Coding Security

Vibe coding security is the practice of controlling risk in apps built by prompting an AI model and shipping what it returns. The flaws are familiar: broken access control, permissive database policies, and exposed keys. The difference is that the review and scanning steps that catch them often sit outside the path. This guide covers risks, incidents, tools, and policy.

Key Takeaways

  • Vibe coding security is mostly a missing-pipeline problem. The access-control and secrets flaws are familiar, but the pull request, CI scan, and change record that catch them are often absent.
  • In the SusVibes benchmark, 57% of one agent’s solutions worked but only 11.8% were secure, and adding vulnerability hints to the prompt did not close the gap.
  • A Supabase public key is designed to be exposed, but the data behind it still needs row-level security. Moltbook and CVE-2025-48757 show what one missing or weak policy hands to an unauthenticated request.
  • Builder scans catch common flaws but, by Lovable’s own account, cannot guarantee complete security. Test each app as two users and as an anonymous caller before it goes public.
  • A workable policy tiers apps by the data they touch, registers them without penalty, and names an owner for each one.

What Is Vibe Coding and Why It Creates a New Security Problem

Vibe coding is building software by describing what you want to an AI model and accepting the code it writes without reading it. Andrej Karpathy named the practice in a February 2, 2025 post: “There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” He added, “I ‘Accept All’ always, I don’t read the diffs anymore.” Collins Dictionary named it Word of the Year for 2025.

That second quote is the security problem. On a browser-based app builder, the prompt, code, database, hosting, and publish button sit in one tab. An app can reach a public URL before anyone opens the diff. Programs focused on security for AI cover models, pipelines, and keys; vibe coding adds the apps those tools produce, expanding the scope to AI AppGen security.

From Accepting the Diff to Shipping the App

The obvious objection is that this is ordinary insecure code, which static analysis already handles. It does, once the code reaches a repository, a pull request, or a CI job, where scanners hook in. A builder that hosts the code and publishes it from a button offers none of the three.

The Biggest Vibe Coding Security Risks for Engineering and Security Teams

The vibe coding security risks split by team. Engineers accept code they cannot explain, and security teams inherit apps they did not know existed.

Risk for the Engineering Team

In SusVibes, a benchmark of 186 real-world feature tasks accepted at ICML 2026, “57% of the solutions from SWE-Agent with Claude 4 Sonnet are functionally correct, only 11.8% are secure.” The authors add that “augmenting the feature request with vulnerability hints, cannot mitigate these security issues,” so a stronger prompt is no substitute for a check.

The second risk is provenance. The Cloud Security Alliance (CSA) notes that a prompt-built app “may exist only as a deployed service with no retrievable source code, no dependency inventory, and no record of the AI model version or prompt that generated it.” Fixing a flaw may mean issuing another prompt, with no guarantee the resulting code is more secure.

Risk for the Security Team

For security, the risk is an app with real users, no inventory entry, no change record, and sometimes a personal account behind it. David Chismon, the UK National Cyber Security Centre’s (NCSC) CTO for Architecture, expects that “Over the next 5 years it will become increasingly common to see AI-written code in production systems that a human has never reviewed or even looked at.” Georgia Tech’s Vibe Security Radar traces published vulnerabilities to AI coding tools: about 18 cases in the second half of 2025, then 56 in the first three months of 2026.

Common Vibe Coding Security Vulnerabilities in AI-Written Code

The vibe coding security vulnerabilities documented in the incidents below fall into one category of the OWASP Top 10. The 2025 edition keeps Broken Access Control at A01 and states the rule these apps break: “Access control is only effective when implemented in trusted server-side code or serverless APIs, where the attacker cannot modify the access control check or metadata.”

Authorization That Lives in the Browser

In a generated app, a hidden admin button can look like access control in a demo. OWASP’s third attack scenario describes the result: “An application puts all of their access control in their front-end.” The attacker skips the page and calls the endpoint with curl, so every check that runs only in JavaScript is optional.

Database Policies That Exist but Do Not Protect

Lovable, for one, pairs a generated front end with a Postgres backend built on Supabase and uses row-level security (RLS) policies to decide who reads which rows. NVD’s record for CVE-2025-48757, marked Disputed, shows what a weak policy allows: “An insufficient database Row-Level Security policy in Lovable through 2025-04-15 allows remote unauthenticated attackers to read or write to arbitrary database tables of generated sites.” Researchers Matt Palmer, then in developer relations at Replit (a Lovable competitor), and Kody Low reported exposed endpoints in 170 of the 1,645 Lovable projects they checked.

Secrets Bundled Into Client Code

A key in the browser is not always a leak. Supabase says its publishable (anon) key is “safe to expose with RLS enabled,” and Palmer wrote that his core finding “was not an exposed public API key as we initially thought” but the absent RLS configuration. The real leaks are private keys. Vite, a common front-end build tool, warns that “VITE_* variables should not contain sensitive information such as API keys” because their values ship in the bundle.

Palmer’s scan also found developer keys returned by open endpoints, including one named get_gemini_api_key. Run secrets detection on any code a builder can export. Invented package names are a separate supply-chain risk, covered under hallucinated dependencies and slopsquatting.

Real-World Vibe Coding Security Issues: What Has Already Gone Wrong

Three 2026 incidents show these vibe coding security issues in production.

The Lovable App That Exposed 18,697 Records

In February 2026, The Register reported on an exam and grading app hosted on Lovable and showcased on its Discover page. Researcher Taimur Khan found 16 vulnerabilities, six critical, including an authentication function the AI wrote with its logic inverted. “The guard blocks the people it should allow and allows the people it should block,” Khan said. By Khan’s count, 18,697 user records were exposed, including 14,928 unique email addresses, 4,538 student accounts, and 870 users with full personal data.

Khan’s broader diagnosis was that Lovable apps break when nobody implements protections such as Supabase’s row-level security and role-based access. Lovable’s CISO, Igor Andriushchenko, said every project gets a free scan before publishing and “Ultimately, it is at the discretion of the user to implement these recommendations.” He added that the vulnerable database was not hosted by Lovable.

A Public Key That Opened a Whole Database

Moltbook, a vibe-coded social network for AI agents, shipped a Supabase API key in client-side JavaScript with no RLS policies behind it. Researchers reported in February 2026 that the key gave unauthenticated read and write access to the production database, including 1.5 million API authentication tokens that let an attacker impersonate any agent. The key was public by design; the missing policy effectively gave it unrestricted database access.

When the Platform Itself Regresses

Lovable’s post-mortem states that “Between February 3, 2026 and April 20, 2026 public project chat history and source code could potentially be accessed by any Lovable user provided they had a project link.” HackerOne triagers closed researchers’ reports from February 22 onward because Lovable’s documentation still called the behavior intended.

Lovable also wrote that “Many users understood ‘public’ to mean their published app was visible on the web, not that someone could read through their build conversation.” Build chats can hold pasted keys, schemas, and sample records, so chat history is sensitive data.

How to Secure Vibe-Coded Apps From Code to Cloud

Even when the traditional pipeline is skipped, three control points remain: the exported code, the publish decision, and what the app reaches in your cloud.

  1. Scan the code wherever it exists. For any app above the prototype tier, require a repository sync or export, then scan it for code flaws, vulnerable dependencies, and secrets. CSA’s governance note says that where export is supported, “automated scanning should be mandatory before an application is deployed to handle regulated data.”
  2. Test behavior before the publish button. Sign in as two users and try to read each other’s records, then call every endpoint with no session. OWASP asks developers and QA staff to “include functional access control in their unit and integration tests.” A dynamic application security testing (DAST) scan of the staging URL can reach endpoints nobody wrote tests for.
  3. Map what the app can reach in the cloud. List the databases, buckets, keys, and identities each app uses, and cut each to least privilege. Alert on anonymous reads of tables holding personal data. AI security posture management (AI-SPM) extends that inventory to models, and secure AI innovation in the cloud covers the wider program.

Best Vibe Coding Security Tools in 2026 (and What to Look For)

Shortlists of the best vibe coding security tools tend to mix three categories. Each sees a different part of the problem, so ask which gap each one closes.

Three Categories on Every Shortlist

CategoryWhere it runsWhat it catchesWhat it misses
Pre-merge scanningThe IDE, coding agent, or pull requestInsecure patterns, vulnerable dependencies, committed secretsApps that never reach a repository
Builder-native scansThe app builder, before publishMissing or permissive access rules, vulnerable dependenciesAnything outside its checks, and whether anyone acts on it
Deployed-app and cloud posture scanningThe live URL and the accounts behind itPublic endpoints, anonymous data access, data and identity reachCode paths it cannot exercise from outside

Palmer wrote in May 2025 that Lovable’s scanner then “merely checks for the existence of any RLS policy, not its correctness or alignment with application logic.” Lovable’s current docs add an optional logic review but warn that its scans “cannot guarantee complete security.” In the grading-app case, Lovable says the scan’s recommendations went unimplemented.

Questions to Ask Before You Buy

  • Does it find apps nobody registered, or only the ones you point it at?
  • Does it test whether a control works, or only that one exists?
  • Does it still work when the builder offers no source export?
  • Can it tie a finding to the data and identities the app reaches?
  • Can it run before publish without blocking prototypes?

For the code-side tools, the guide to best AI code security solutions compares vendors in depth.

Building Guardrails: A Vibe Coding Security Policy for Your Organization

CSA’s June 2026 note advises “establishing visibility before attempting control,” with a lightweight registration process and “non-punitive framing.”

  1. Scope. Cover browser app builders, coding agents, and AI features inside sanctioned low-code platforms, where, in CSA’s words, “The platform is sanctioned; the application is not.”
  2. Approved tools. Keep a short list of approved builders and agents, and require company identities and single sign-on. Anything else counts as unsanctioned AI tools and shadow AI and goes through intake.
  3. Data classes allowed. Follow CSA’s three tiers: internal tools with non-sensitive data, apps with personal or regulated data, and apps that call AI APIs or act as agents. The NCSC’s spectrum guidance puts authentication logic and “handling secrets, tokens, or credentials” at the strict end.
  4. Publish approval. Tier 1 apps self-certify. Tier 2 and 3 apps need the two-user and anonymous tests above and a named reviewer before going public.
  5. Ownership. Record each app’s builder, account, data reach, and owner, and transfer it when the builder leaves.
  6. Review cadence. Re-test tier 2 and 3 apps after each major re-prompt and at least quarterly, since a re-prompt can rewrite access logic.

How Orca Connects Vibe Coding Security to Live Cloud Exposure

When the traditional pipeline is absent, security controls can still operate at the deployed app and cloud-account level. That is where Orca adds visibility. Orca extends the agentless discovery it already uses to detect shadow AI across the cloud estate to AI app generator platforms, so apps surface without anyone registering them. For each app, Orca “identifies the misconfigurations they ship with by default, and connects each finding to the surrounding cloud environment.”

Context sets the priority. A permissive policy on an app with no users may be low priority. On an internet-facing app holding student records, the same issue can represent an immediate exposure. Orca evaluates each app by “checking whether it’s internet-facing, whether it touches sensitive data, and how it connects to identities and attack paths.”

AI AppGen Security, announced in July 2026 for general availability later in 2026, is a capability within Orca’s existing AI Security. Its findings are prioritized in the same unified risk model as the rest of the cloud environment.

Get a Demo

Frequently Asked Questions about Vibe Coding Security

Who coined the term vibe coding?

Andrej Karpathy, in a post on February 2, 2025, described “a new kind of coding” in which he accepted every change without reading the diffs. He also set its limit: “It’s not too bad for throwaway weekend projects.”

Is vibe coding safe for internal tools?

Often. NCSC lists “working on internal tools with limited exposure” and cases where “no sensitive data or security functions are involved” as reasons to slide toward full vibe coding. An internal tool that reads HR or customer records fails that test.

Who is responsible when a vibe-coded app leaks data?

Responsibility depends on where the failure occurs, but organizations remain responsible for securing the applications and data they control. NVD records Lovable’s dispute of CVE-2025-48757: “each individual customer of the Lovable platform accepts a responsibility over protecting the data of their application.” Platforms still answer for their own infrastructure, as Lovable’s 2026 regression showed, so contracts should state where that line sits.

Do vibe-coded apps fall under GDPR and HIPAA?

Yes, whenever the data is in scope. CSA notes that these regimes impose obligations “based on the categories of data their applications process, not on how those applications were built.” A vibe-coded app handling protected health information is subject to the same applicable HIPAA requirements as conventionally developed software. An organization that cannot list the app cannot meet GDPR’s accountability principle.

Do AI coding agents in an IDE carry the same risk as app builders?

Less, while an engineer stays in the loop. CSA draws the line at review: trained engineers typically review generated code inside an established workflow, while citizen developers on prompt-to-app platforms often cannot. When agent output lands in a repository, your existing scanners can also apply.