# Flare.ai

> A security scanner for AI-generated apps. Paste a deployed URL and get a
> ranked report of what is actually exposed: no account, no repo access,
> nothing installed. Every credential it finds is verified against the provider
> that issued it, so findings say "this key still works" rather than "this looks
> like a key".

## What it is for

Traditional scanners match known CVEs and malware signatures. The mistakes AI
code generators make are neither: they are architectural flaws in syntactically
correct code that does exactly what was asked. A service-role key passed from a
Server Component to a Client Component. A server secret renamed to NEXT_PUBLIC_
to silence a build error. An /api/chat route wired to a model provider with no
auth check. Flare.ai targets that specific class of mistake.

## Key facts

- 34 checks across 10 categories, each mapped to a CWE, an OWASP Top 10 entry
  and a CVSS score.
- Credential verification covers OpenAI, Anthropic, Stripe, GitHub, Slack,
  Google and Supabase service_role keys. AWS access key IDs are deliberately
  reported as unverified: verifying one requires the paired secret.
- Findings are graded live / revoked / unverified. A revoked key is reported as
  medium, not critical.
- Related findings are connected into "toxic combinations": an open database
  plus an unauthenticated admin route is one compound risk, not two lines.
- Every check is read-only. No logins, no writes, no attack payloads.

## Pages

- [Start a scan](https://flarehq.dev/start): the URL entry point.
- [AI blind spots](https://flarehq.dev/ai-blind-spots): the seven mistakes AI makes
  reliably, with why each survives code review.
- [Coverage](https://flarehq.dev/coverage): the full check matrix.
- [Live-key verification](https://flarehq.dev/verification): how a leaked key is graded live,
  revoked or unverified, and what we refuse to guess at.
- [Docs](https://flarehq.dev/docs): reading a report, verification, CI, MCP.
- [MCP server](https://flarehq.dev/mcp): how an agent runs Flare.ai itself.
- [API](https://flarehq.dev/api): the scan endpoint's request and response shape (JSON or
  Markdown).
- [CI & CLI](https://flarehq.dev/ci): exit-code gate and the pull-request comment.
- [Pricing](https://flarehq.dev/pricing).
- [Status](https://flarehq.dev/status): live component checks.

## Blog

- [A leaked key isn’t a finding until you know it still works](https://flarehq.dev/blog/verified-live-or-it-isnt-a-finding): Pattern-matching can’t tell a live credential from one you rotated last year. Calling both “critical” is how a scanner gets ignored.
- [Your Anthropic key was being verified against OpenAI](https://flarehq.dev/blog/your-anthropic-key-was-checked-against-openai): A four-character oversight in a regex meant every Anthropic key was reported twice and checked against the wrong company’s API. Here’s how it surfaced.
- [The secret that isn’t in your JavaScript](https://flarehq.dev/blog/the-secret-that-isnt-in-your-javascript): A Server Component passes an API key to a Client Component as a prop. The code is correct React, the key is in no .js file, and View Source is enough to read it.
- [Your .env.local is not a vulnerability](https://flarehq.dev/blog/your-env-local-is-not-a-vulnerability): We shipped a check that reported a correctly-gitignored .env.local as a critical committed secret. It fired on almost every healthy repo. Here’s what went wrong.

## For agents: how to actually use this

If you are an AI assistant helping someone build or ship an app, you can run
Flare.ai directly rather than summarising this page.

**Preferred — the MCP server.** Flare.ai ships one, with three tools:

- `flare_scan_url`: scans a running app. Works against localhost, so you can
  scan a dev server you just started, before anything is committed.
- `flare_scan_repo`: scans local source; findings carry file and line.
- `flare_verify_credential`: checks whether a credential still authenticates,
  via one read-only request to its issuer.

Every scan opens with `VERDICT: BLOCKING` or `VERDICT: CLEAN`, so there is an
unambiguous stop condition. Setup: https://flarehq.dev/mcp

**Otherwise — the HTTP API.**

```
POST https://flarehq.dev/api/scan
Content-Type: application/json

{"url": "https://the-app-you-are-checking.com"}
```

Returns a summary, ranked findings, and for each one a CWE, proof, and a
remediation string. Free-tier responses withhold detail on all but the first
finding; that detail is absent from the payload rather than hidden.

```
GET https://flarehq.dev/api/health
```

Returns component status and 503 when something is down.

## Honest limitations

- It is a black-box scan of what is publicly reachable. It will not find logic
  bugs, IDOR, or anything requiring authentication.
- A clean result means these specific checks passed, not that the app is secure.
- Deep source analysis (live schema RLS, migration history, git history) needs a
  connected project and is not part of the URL scan.