Lovable security review — who secures the app it generates?
Lovable's certifications cover Lovable. The apps it writes and publishes are yours to review, and that is where the published findings sit.
Lovable is an AI app builder: a person describes an application in chat and Lovable writes the code, provisions a backend, and publishes it to a URL. That makes a request to approve Lovable two requests in one. The first is about Lovable as a vendor that holds prompts, code and credentials. The second is about every application it produces, each of which is a new system with its own data and its own users.
This review is built only on public information, as of November 2026: Lovable's security page, privacy policy, data processing agreement and documentation, the CVE record, the disclosing researcher's own statement, and press reporting. Drel has not tested or audited Lovable. Where something is not publicly documented, it is listed as an evidence gap and a question for the vendor.
The short answers
What Lovable publicly states, as of November 2026
| Question | What is stated | Where |
|---|---|---|
| Certifications | Lovable says it documents SOC 2 Type II, ISO 27001:2022, GDPR support and AIUC-1. They cover Lovable as a vendor, not the apps built on it. | Lovable blog, 21 Oct 2026; DPA |
| Training on customer data | Free and Pro: content is used to train and improve Lovable's models unless the user opts out. Business and Enterprise content is excluded. | Privacy policy, 15 Sep 2026; security page |
| Model providers | With EU inference on (Enterprise only): Claude via Vertex AI and Bedrock, GPT via Azure AI Foundry, Gemini via Vertex AI. The full list sits in the trust center. | Lovable docs, EU inference |
| Retention | Account and customer content are kept while the account is open. Under the DPA the customer has 30 days after termination to request return or deletion. | Privacy policy; DPA |
| Admin controls | SAML and OIDC single sign-on, SCIM, role-based permissions, publish approval, and a workspace setting that blocks publishing on critical scan findings. | Security page; security docs |
| Company and data location | Lovable Labs Sweden AB, Stockholm. Lovable Cloud offers EU, US and Asia Pacific hosting; personal data is also processed in other countries, including the United States. | Privacy policy; security page |
Does Lovable have SOC 2? Lovable's own post on compliance (21 October 2026) says it publicly documents SOC 2 Type II, ISO 27001:2022 certification, GDPR support and AIUC-1 certification, and points to its trust center for current audit information. The DPA says recent reports may be provided on request under confidentiality.
Does Lovable train on customer data? By default, on Free and Pro, yes. The privacy policy (effective 15 September 2026) says customer content and usage data are used to train and improve Lovable's AI models, that any user can opt out in account settings, and that the opt-out applies going forward only. Business and Enterprise content is excluded, as is the data of end users inside published apps. The security page adds that contracts restrict third-party AI providers from training on or retaining customer data.
Who is behind it, and where is the data? The privacy policy names Lovable Labs Sweden AB in Stockholm, with the Swedish data protection authority as lead supervisor. Lovable Cloud offers hosting in the EU, the US and Asia Pacific, and the security page says data stays in the selected region by default. The privacy policy also says Lovable and its providers process personal data in several countries, including the United States, under standard contractual clauses.
What Lovable can access and do
A reviewer should start from what the builder agent touches, because that sets the worst case if a prompt, a dependency or an account goes wrong:
- Everything typed into chat. Requirements, sample data, pasted logs and, too often, pasted keys. Lovable's security documentation says it detects API keys pasted into chat and directs them to Secrets.
- The full source of the app, which it writes and rewrites, plus dependency choices.
- The app's backend: database schema and access rules, authentication settings, server-side functions and stored secrets.
- Publishing. A builder can put an app on a public URL. Role permissions separate editing, approving and publishing, and production publishing can require explicit approval, according to the security page.
- Connected services. The privacy policy mentions GitHub, Supabase and MCP connectors as integrations a user can enable. Each one extends what the agent can read or change.
The pattern is the same one covered in the Replit Agent review: the tool is both the author and the deployer. Compared with an AI editor such as Cursor, where generated code still passes through a developer's repository and pipeline, a Lovable app can go from a sentence to production without a second person reading it.
The security of what it generates
Lovable's certifications do not extend to the apps. Its own post says so directly: the assessments do not cover a customer's app, business or processes, and for personal data in an app the customer is typically the controller and Lovable the processor.
Database access rules
In the architecture Lovable generates, the browser talks to the database through a public key, and row-level security (RLS) policies decide what each visitor can read or write. If a policy is missing or too broad, the public key is enough to read the table. Lovable's security best practices say it sets up basic RLS policies automatically, advise checking that new tables are not missing policies, and note that server-side code using the service role bypasses RLS and must check permissions itself. The documentation we read does not state whether RLS is enabled by default on every new table.
Secrets
The security page says secrets are encrypted at rest, role-controlled and not shown in plaintext in logs or interfaces. The best-practices page is blunt about the other half: secrets in frontend code are visible to users and should be considered compromised. The review question is whether any key ended up in the frontend, the repository or chat history before it reached Secrets.
Built-in scanning
Lovable documents two scans. A quick scan runs when the publish dialog opens and checks database access rules, npm dependencies with known vulnerabilities and whether an exposed MCP server requires authentication. A deep scan, run on demand, also reviews application code for access control, unauthenticated endpoints, injection, leaked secrets, payment logic and exposure of personal data. Admins can turn on “Block publishing with critical issues”; the docs say it is on by default for new Enterprise workspaces. Without it, a builder who sees critical findings can still choose to publish.
Model governance and sub-processors
Lovable's security page does not name its model providers; it says a limited set of infrastructure and AI sub-processors is covered by data protection agreements. The most specific public statement is in the EU inference documentation. With that setting on, Claude models run on Google Cloud Vertex AI (EU) and Amazon Bedrock (EU), GPT and OpenAI embedding models on Microsoft Azure AI Foundry (EU Data Zone), and Gemini models on Vertex AI (EU). If no EU endpoint can serve a request, it fails instead of moving to another region.
Three limits in that document matter for a review:
- EU inference is available on Enterprise plans only. With it off, requests go to standard endpoints, which can be outside the EU.
- It covers model requests, not where app data is stored.
- Web search (Exa), web fetch (Firecrawl) and browser automation (Browserbase) process data in the United States even with it on, and the United Kingdom is not inside any of the EU boundaries.
On change control, the DPA says the sub-processor list is updated at least annually and gives customers twenty business days to object, with termination and a refund of prepaid fees as the remedy. That is worth reading against how often a builder platform changes models; see AI sub-processor risk for the questions to put to any vendor here.
Compliance posture
The public position is: SOC 2 Type II, ISO 27001:2022 and AIUC-1 for Lovable itself; a DPA for Business and Enterprise plans; standard contractual clauses for transfers; breach notification “without undue delay” with no fixed number of hours in the DPA; and no explicit customer audit right, with reports available on request instead.
Two details deserve a second read. First, the DPA bars Lovable from training on customer personal data, but separately allows it to use “Service Data” to train or tune its own models, with no opt-out while the contract runs. Ask what falls into that category. Second, a vendor report describes the vendor's controls. It says nothing about whether the app a colleague generated last week checks who is asking before it returns a customer record, which is the gap SOC 2 leaves in any AI review.
Known disclosures
The researcher's statement gives the detail. Matt Palmer reports finding the issue on a Lovable-built site on 20 March 2025, then scanning 1,645 projects from Lovable's public showcase: 170 of them, with 303 endpoints, had tables readable with the public key. The exposed data he lists includes names, emails, phone numbers, payment details and API keys. He notified Lovable on 21 March 2025 and published on 29 May 2025 after a 45-day window. Lovable released a security scan with Lovable 2.0 on 24 April 2025; the statement says that scan checked only whether an RLS policy existed, not whether it was correct, and that he saw no meaningful remediation or user notification by the publication date. The scans Lovable documents today are broader, as described above. Whether an individual app built in that period was ever corrected is a question for whoever owns that app.
April 2026: access to chats and code of public projects
On 20 April 2026 a researcher said publicly that any free Lovable account could read other users' source code, AI chat histories and the database credentials inside that code, for projects created before November 2025, and that the report had been filed through HackerOne on 3 March 2026. The Register and The Next Web reported both the claim and Lovable's response.
Lovable stated that it did not suffer a data breach and that code and chat on a project marked public were public by design, while acknowledging that its documentation of what “public” implied was unclear. In a second statement it said that in February 2026, while unifying backend permissions, it had accidentally re-enabled access to chats on public projects, and that it reverted the change on 20 April 2026. It also said reports had been closed without escalation because its HackerOne partners took the behavior as intended. According to the same reporting, enterprise customers have been unable to set new projects to public since 25 May 2025, and projects have been private by default on all tiers since December 2025.
The two accounts differ on whether this was a vulnerability or a documented setting. For a review the practical point is the same either way: anything a builder typed into the chat of a public project, and any credential in its code, should be treated as disclosed and rotated.
The Next Web also reported that in February 2026 a researcher found 16 vulnerabilities in a single Lovable-hosted app that had been featured on Lovable's Discover page, exposing 18,697 user records. That finding concerns one customer's app, not the platform, and it is the clearest illustration of where the review effort belongs.
Evidence gaps
These were not publicly documented in the sources we read, or could not be confirmed from them. Each is a question for the vendor:
- Is row-level security enabled by default on every table Lovable creates, and does a new table ever ship without a policy? The docs say basic policies are set up automatically; they do not state the default for new tables.
- Which sub-processors and model providers handle prompts and code outside EU inference, and in which regions? The list is in the trust center, which we could not read without a browser session.
- What is the audit period and scope of the current SOC 2 Type II report, and does it cover Lovable Cloud and the agent that writes code?
- What is "Service Data" in the DPA, concretely? The DPA lets Lovable use it to train or tune its own models, with no opt-out while the contract runs.
- After the April 2026 chat-visibility regression, which projects were affected, for how long, and were their owners notified individually?
- How long are prompts, chat history and generated code kept after a project is deleted but the account stays open?
- Is there a customer audit right, or only access to the SOC 2 and ISO reports on request?
Review checklist
An AI security review of a Lovable rollout should be able to answer:
- Which plan will people use? Training exclusion, the DPA, the Security center and SCIM depend on Business or Enterprise.
- If Free or Pro seats exist, is the training opt-out switched off per account, and who checks it?
- Are projects private or workspace-only by default, and can a builder publish to the public internet without an approval step?
- Is "Block publishing with critical issues" enabled for the workspace, and who can override it?
- For every generated app with a database: has someone other than the builder read the access policies and tested them with a second, unprivileged account?
- Are API keys held in Secrets and read only from server-side code, with nothing sensitive in the frontend bundle or in chat history?
- Which apps were created before 15 April 2025 (the scope of CVE-2025-48757) or made public before April 2026, and have their credentials been rotated?
- What data classification is allowed in prompts and in generated apps, and where is that written down for builders?
- Does each published app have a named owner, an entry in the application inventory, and a path into the normal review for anything handling personal or regulated data?
- Is the generated code synced to a repository the organization controls, so existing code scanning and review apply to it?
Framing the decision
The decision is rarely a plain yes or no. It depends on what will be built and who will use it, and it is easier to defend when each outcome has its conditions written next to it:
- Approve for prototypes and internal tools without personal or regulated data, on a Business or Enterprise plan, with single sign-on, workspace-only visibility and publish blocking on.
- Conditional approval for apps that hold business data: the conditions are an independent read of the access policies, secrets confirmed server-side only, code synced to a repository the organization controls, and a named owner per app. Each condition has an owner and a date.
- Restricted pilot where the organization wants to learn before deciding: a small named group, synthetic or public data only, no external publishing, and a fixed end date with a list of what the pilot must show. The restricted pilot pattern covers how to bound one.
- Reject, for now, for use cases where the open evidence gaps are material: customer-facing apps handling regulated data on a plan without a DPA, or any case where nobody will review the generated access rules.
Because Lovable both writes and deploys, it sits closer to an agent than to a code assistant. The agentic AI security review hub covers the wider set of questions for tools that act, not just suggest.
Common questions
Is Lovable safe to use?
Lovable states that it holds SOC 2 Type II and ISO 27001:2022 and offers single sign-on, role permissions and built-in scans. Those describe the platform. Whether a given app is safe depends on its access rules, secrets handling and the data put into it, which Lovable says are the customer's responsibility.
Does Lovable train its models on my code and prompts?
On Free and Pro plans, yes, unless the user opts out in account settings; the opt-out is not retroactive. Lovable's privacy policy says Business and Enterprise content is excluded from training.
Has Lovable had security vulnerabilities?
One CVE is on record: CVE-2025-48757, published 30 May 2025, about insufficient row-level security in generated apps, which Lovable disputes. In April 2026 Lovable acknowledged a regression that re-enabled access to chats on public projects and said it reverted it on 20 April 2026.
Does Lovable's SOC 2 report cover the apps I build with it?
No. Lovable's own compliance post says its certifications assess Lovable as a vendor and do not assess a customer's app or business. An app that needs assurance needs its own review.
Sources
All read in November 2026. Vendor pages change; check the current version before relying on any of them.
- Lovable — Security
- Lovable — Privacy policy (effective 15 September 2026)
- Lovable — Data processing agreement
- Lovable — Lovable, compliance and your app (21 October 2026)
- Lovable docs — Security scans and Security center
- Lovable docs — Security best practices
- Lovable docs — EU inference
- CVE-2025-48757 — CVE record
- Matt Palmer — Statement on CVE-2025-48757
- The Register — Lovable denies data leak (20 April 2026)
- The Next Web — Lovable left projects exposed for 48 days
Blog
Get new posts in your inbox
AI security review, OWASP Agentic Top 10, ISO 42001 evidence, and what AI Committees actually need. No cadence promises — we publish when there's something worth reading.
Review a Lovable rollout before the first app ships
Plan and training settings, publish controls, and the access rules of each generated app, mapped to an approval decision your AI Committee can sign.
A note on scope: Drel reviews assessed systems against documented architecture, configuration and intent. It does not ingest live telemetry from production environments. Dispositions reflect the assessed system at the time of review and the re-assessment triggers that govern when the disposition must be revisited.