Do I need insurance for an app I built with AI (vibe coding)?

What “vibe-coded” changes about liability, and what it does not

Vibe coding means describing what you want to an AI tool, such as Cursor, Lovable, Bolt, Replit, v0 or Claude Code, and shipping what it writes, often without reading all of it. It is a way of building, not a legal category. A customer who loses money because your app failed does not ask how it was written. They ask you.

The tools say the same thing in their own terms. Replit’s terms warn that code its AI generates or suggests may be erroneous or incomplete, and Lovable’s supplier position on its published vulnerability record is that each customer is responsible for protecting the data in their own application. Read that plainly: the tool is a contractor you cannot hold to the result, and the liability sits with you.

What does change is the odds and the questions. The documented failures below are almost all the same few mistakes, repeated, and an insurer who knows that will ask whether anyone with security experience looked at the app before it held other people’s data.

The failures that have actually been documented

The most common one is an unlocked database. Supabase, the database behind many AI-built apps, gives the browser a public key and relies on row level security rules to decide who can read which rows. When those rules are missing, the public key opens everything. Wiz found exactly this at Moltbook in early 2026, with the founder saying he wrote none of the code: about 1.5 million authentication tokens, about 35,000 email addresses and private messages, some containing other companies’ API keys. A separate CVE record describes the same class of problem across sites generated by Lovable, through April 2025, and the supplier disputes who is responsible for it.

The second is a platform door left open. In July 2025 Wiz showed that Base44, owned by Wix, let an outsider create a verified account on a private app knowing only its app ID. It was fixed within a day and the vendor found no customer impact, but an app built on a shared platform inherits the platform’s mistakes, and the app’s owner still has to tell the users if their data is touched.

The third is the agent that acts. In July 2025 an investor testing Replit’s agent reported it deleted his production database during a code freeze. By his own account the data was recovered and the app was an experiment, so it is not a loss story. It is a warning that an AI agent with write access to production can do damage nobody approved, which is a business-interruption and data-loss question, not only a coding one.

Underneath all three is the wider pattern. Veracode’s 2025 testing found models picking the insecure way to write code 45 percent of the time when both ways were available, and that it did not improve as models got better at writing working code. None of this means AI-built apps cannot be run safely. It means the review step you skip is the step that decides the claim.

Which policy answers which failure

Technology E&O, short for errors and omissions, is the policy for a customer who says your app did not work and it cost them money: a bug that corrupts their data, an outage in the middle of their business day, a feature that did what the screen said not to. It pays to defend you and to settle covered claims. Most apps selling to businesses need it because the customer contract asks for it.

Cyber liability is the policy for the breach. First-party cyber pays your own costs of responding, such as forensic investigation, notifying users, credit monitoring and restoring systems. Third-party cyber pays when users or customers sue or regulators ask questions after their data leaked. The two are often sold together and the split matters, so the guide on first-party and third-party cyber explains it.

Media and intellectual property coverage answers a claim that your app, or what it showed, copied someone else’s material. Generative tools can produce code, text and images that resemble existing work, and not every policy treats that the same way.

General liability is the policy people already own, and it is the least likely to help. It is written for accidental injury and property damage, not for lost data or failed software, and a newer family of endorsements lets insurers exclude generative AI outright. That is its own page, and this one links to it rather than repeating it.

Contingent bodily injury sits on top of all of these if your app tells people what to do with their bodies, their cars or their equipment. It is the grant that lets E&O answer an injury caused by your app’s output. A fitness, diet or health app needs it; a scheduling tool usually does not.

Ownership, platform rules and what you can promise

Who owns the code is less settled than people assume. The U.S. Copyright Office’s January 2025 report said that prompts alone do not make you the author of what the model writes, though your own edits, selection and arrangement can be protected. For a founder this matters in two places: a buyer who wants to acquire your app will ask what you actually own, and a customer contract that has you warrant ownership or non-infringement is making you promise something you may not be able to prove.

Platform rules matter next. Apple requires a privacy policy, plain disclosure of what you collect and permission before sharing personal data with outside services, including third-party AI. If your app sends a user’s information to an AI model and your policy does not say so, you have made a promise you are breaking, and that is the kind of claim a regulator or a class action starts from.

Breach notice is the part that surprises people. Every state has a law requiring notice when personal information is exposed, and the duty falls on the business that holds the data, not the tool that built it. Notice, forensics and credit monitoring are real costs even when nobody sues, which is exactly what first-party cyber is for.

An example from our own desk

Our founder, Kevin Kelsey, builds apps himself. TrainRec, at www.trainrec.app, is a workout and food app that runs training plans made with your own AI assistant, guides each workout one set at a time, logs food from a description and syncs with Apple Health. It carries the exposures on this page, and its footer says it is not medical advice. We mention it so you know the person writing this has had to answer the same questions for his own product, not to suggest any particular coverage fits yours.