San Francisco – November 06, 2026 -- Constructive has released Safegres, an open-source verifier that grades which roles can reach which rows in a live PostgreSQL database, after recent scans found 24,664 of 159,041 Lovable-built projects serving records to anonymous visitors.
Safegres audits row-level security (RLS) policies and grants in read-only mode. It never touches application data or changes anything, so teams can run it against a production copy or automatically on every database change.
Scans found thousands of AI-built apps exposing Postgres data
On Nov. 2, Hamburg security firm Crissy Field and the German newspaper Handelsblatt scanned 159,041 projects built with AI app builder Lovable. They found 24,664 sites whose Postgres databases served records to anonymous visitors, 7,403 of them exposing email addresses.
Other research points the same way. UpGuard found 16,326 databases exposing readable tables across about 300,000 domains. Shippable found public tables in 25.6% of 4,994 Bolt.new projects. CVE-2025-48757 covered 303 exposed endpoints across 170 Lovable apps.
Supabase stopped automatically granting access to new tables on Oct. 30, 2026. Grants decide whether a role can reach a table, and RLS decides which rows it sees. Agents now create 80% of databases on Neon, according to Databricks, and 70% of new Supabase databases, according to Supabase.
Safegres runs 30+ security and 10+ performance checks on every change
The two scores are graded separately. Safegres follows permissions through inherited roles, shared defaults, views and functions to show what each role can actually reach.
Results are deterministic: no AI judge is involved, the same database always gets the same grade, and each report is fingerprinted. Constructive validates the tool against a library of sample databases with known flaws, checking that it finds what it should and nothing it shouldn't.
In CI/CD, including on GitHub pull requests, Safegres flags only what a given change introduces. Ready-made settings cover Supabase, PostgREST, Hasura, Graphile and Constructive.
Parser downloaded more than 9 million times a month underpins the tool
Safegres is built on Constructive's open-source toolkit, which translates SQL into a structured form programs can read, check and rewrite using PostgreSQL's own grammar. Constructive's Postgres parser is downloaded more than 9 million times a month. Its parser tools have been used in more than 1,100 public GitHub projects, including projects from teams at Canva, Gel Data (acquired by Vercel), Neon and Supabase.
Dan Lynch, Founder and CEO of Constructive, said that with security incidents piling up across AI-built apps, these checks "can't be optional." Co-founder Anmol Yadav, Head of Infrastructure, said Safegres is meant to run on every commit whether a person or an agent wrote the change.
Sample databases double as a test bed for coding agents
Teams can give an agent a flawed database, then re-audit to see whether it fixed the problem without creating new ones. Lynch said coding agents that can see Safegres results in CI and GitHub checks pick up changes in the security and performance scores and fix issues on their own. He described the verifier as a feedback loop for building evals and self-improving agents.
Safegres is released under the MIT license and works with Supabase projects, any PostgreSQL database and common CI pipelines. It joins two other Constructive open-source tools: pgpm, which organizes Postgres code into packages, and pgsql-test, which runs tests against a real database.