Lovable Ships Real Scans as Agent Attack Surface Grows

Two things happened at once this week. Lovable added real security tooling — scans that check row-level security and exposed endpoints, plus a dashboard to track findings across projects. At the same time, reports describe coding agents getting fooled by a single fake error message. Both are worth understanding if you shipped an app with an AI builder.

TL;DR

• Lovable's new Basic and Deep scan profiles check RLS policies, access control, exposed secrets, and unprotected backend endpoints — not just "is a key visible."
• Workspace insights now gives Enterprise admins a security and PII review priority per project.
• Eleven new Lovable app connectors (Chargebee, Pipedrive, WooCommerce, and others) mean more third-party permissions to check before you connect them.
• Vibe Coding Weekly #36 reports a fake Sentry error message was enough to hijack major coding agents — a reminder these tools trust what they read.

Lovable is scanning for the risks that actually matter

Lovable's new scan profiles are a good sign for anyone building on the platform. Basic scan checks row-level security policies, database schema, access control, and dependency vulnerabilities — and it now runs automatically when you open the publish dialog. Deep scan goes further: agentic review for unprotected backend endpoints, exposed secrets, and unsafe input handling.

This matters because the common failure mode in AI-built apps was never a public key. Supabase anon keys and Firebase web configs are meant to be public. The real risk is a database table with RLS off, or an endpoint anyone can hit without a check. Lovable's scan is aimed squarely at that gap. If you're on Lovable, run Deep scan before your next publish and read what it flags — don't just check that it passed.

Enterprise teams also got Workspace insights: one place to see security findings, PII findings, ownership, and a review priority (High, Medium, Low) across every project in the workspace. If you're managing more than one Lovable project, this is the first place to look each week.

More connectors, more permissions to check

Lovable added eleven app connectors this month — Chargebee, GatewayAPI, Lexware, Pipedrive, PrestaShop, Sevdesk, Wave, Wix, WooCommerce, Zoho Books, Zoho CRM — plus AWS Athena and Replicate. Each one gives your app new access: billing data, customer records, or a live SQL connection to S3.

None of this is a vulnerability by itself. But every connector is a new set of credentials and permissions living in your project. Before you connect one, check what scopes it asks for and whether your app actually needs write access, or just read.

Separately, the Lovable MCP server is now available on all plans, including Free — and third-party MCP client access is on by default for Free and Pro workspaces. That means tools like Claude, Cursor, or VS Code can create, edit, and deploy your project through natural language, by default. Business and Enterprise admins can turn this off in Settings. Everyone else should at least know it's on.

Agents can be tricked by what they read

Vibe Coding Weekly #36 describes a single fake Sentry error message hijacking major coding agents. The underlying issue: agents that read logs, error messages, or web content as part of their workflow can be manipulated by anything planted in that content, not just by the developer's own prompts. Vibe Coding Weekly's broader read this week is that AI dev tooling is starting to "feel like regulated infrastructure" — governance features like Lovable's are part of that shift.

The practical takeaway: if your build agent pulls in external content — error logs, scraped pages, API responses — treat that content as untrusted input, the same way you'd treat user input in your app.

FAQ

Is a public Supabase or Firebase key the problem here?

No. Anon keys and web configs are meant to be public. The real risk is what they can access — specifically, missing or misconfigured row-level security. That's exactly what Lovable's new Basic scan checks.

Do I need Enterprise to get these security scans?

Basic and Deep scan are described as built into Lovable generally, run from the project's Security view or automatically at publish. Workspace insights, the cross-project governance dashboard, is Enterprise-specific.

Should I turn off third-party MCP access?

If you're on a Free or Pro Lovable workspace, it's on by default. If you don't have a reason for an external tool like Cursor or Claude to edit your project directly, there's no downside to reviewing that setting.

The bottom line

The platforms are catching up. Lovable's scans now check the things that actually cause breaches — RLS, exposed endpoints, unsafe input — not just whether a key is visible. That's real progress. But the ecosystem around these tools is also growing faster: more connectors, more default agent access, and agents that can be fooled by a planted error message. Use the new scan tools. Review what's connected to your project. Don't assume more integrations mean more safety.