Security of AI-generated code: 56% success, and what you can check

Software that launches and performs its requested service can still have an open door. Veracode has measured large models' generated-code security for two years, and the rate has not moved. What the figure covers, and four requirements before publishing.

AI-generated code security plateaus at a 56% test pass rate versus 55% a year earlier, according to Veracode's 2026 GenAI Code Security report published July 28, 2026. The same report says 44% of generation tasks produce a vulnerability.

What the figure covers

Syntax correctness is a solved problem: models write code that compiles. Veracode gives generation tasks to models, then security-tests the results. Code can compile, execute and deliver the requested service while failing that test. Rates by model category are tightly grouped: 51% for coding specialists, 52% for general-purpose models, 56% for reasoning models, 53% for large models against 51% for medium and small ones. Choosing a larger model therefore does not solve the question, and the best tested model stops at 68%. The one-point stagnation over a year is the report's striking fact: models improved on almost everything except this.

Vulnerabilities where the rate collapses

The average hides a sharp gap by vulnerability class. For SQL injection, the attack diverting a query to read an entire database, generated code passes in 83% of cases; for cryptographic algorithms, 87%. For cross-site scripting, allowing a visitor to inject content executed for other visitors, the rate falls to 15%. For log injection, 12%. These are the web's most ordinary vulnerabilities, tried first by attackers, and the ones generation most often misses.

What a competitor does better than us here

Lovable announced on August 5, 2026 that each application published on its platform receives a dedicated security page hosted at the application's address, displaying active controls. ‘Is my application secure?’ becomes something you can consult rather than an intuition. This is a useful admission: security had been invisible to the person commissioning software. Maestro offers no equivalent today, and this is a requirement to put to every tool, ours included.

What automated verification catches, and misses

At Maestro, agent Félix verifies every built stage, and tests are written before code. These tests concern expected behaviour: a quote is created, a total rounds correctly, an unauthorised user cannot see the page. They do not replace security analysis, and we will not write otherwise. Software passing every test may still allow injected content in a comment field because no functional test thinks to paste code into a text area. Functional and security checks are two professions; one does not remove the need for the other.

Four requirements to set

An opinion piece by Yves Wattel (Delinea), published by Journal du Net on August 12, 2026, lists four safeguards against offensive uses of code generation: least privilege, credential rotation, temporary access and behavioural monitoring. Translated for an owner, these become four sentences for whoever builds the software. Every account accesses only what it needs, including yours. Keys and passwords are changed, with the change planned somewhere rather than left to someone's memory. Provider access expires automatically. A record exists of what was consulted and by whom. Requesting these four things at the outset costs nothing; adding them after launch costs a project.

In practice, before publishing

Start with the shortest check: can your database be reached from outside, and has it been reachable since day one? Then review every free-text field, comments, names, messages, and ask what happens if someone pastes something other than text. These checks fit into an evening and cover the two categories where Veracode's measured rate is lowest. The rest belongs to AI development traps and, now with a date attached, what your software must account for.