Application security for AI-built software

Your app works. That's not the same as safe.

An AI-built app can pass every test you'd run and still let a stranger read your users' data. Nothing looks broken — that's why it goes unnoticed. Meshbound finds those holes by rule, fixes them from a library human engineers wrote, and proves each fix holds before you approve it. Not a warning. The fix, and the evidence.

Your AI writes code that works. It doesn't write code that refuses.

01 / the difference Everyone flags.
We fix, and we prove it.

The usual answer.

A warning, and a prompt.

You get a list of warnings and, at best, a prompt to paste back into the same AI that wrote the hole. No way to know whether the patch worked, or what else it changed. The alternative is a three-thousand-dollar audit that's stale by your next push.

Meshbound

A fix, and the evidence.

Meshbound finds holes by rule, not by opinion. Each fix comes from a library written and reviewed by human engineers — not generated on the spot. Before you ever see it, the fix is rehearsed against a throwaway copy of your app: the attack must fail, and a real user must still get their own data. Only then does it reach you, as a change you review and merge. Nothing is auto-applied. Nothing asks for trust it hasn't earned.

verification record Hole closed Access intact
crafted attack — foreign user id

GET /rest/v1/orders?user_id=eq.a71f
→ 200 OK · 0 rows

legitimate request — own session

GET /rest/v1/orders
→ 200 OK · 14 rows, all their own

Every fix ships with this record. That proof is the product.

02 / the workspace Findings, fixes, and the proof — in one ledger.

meshbound.dev/workspace/tastelog
tastelog · main scan #147 · 12 min ago

Every finding carries its own history: what was detected, what was attempted, what was refused, and why.

03 / how it works A guardian, not an audit.

An audit tells you what was true on Tuesday. Meshbound connects once and re-checks on every push — because the same prompt that adds a feature can quietly reopen the hole you closed last week. The problem doesn't hold still, so neither does the check.

04 / what it catches The mistakes AI makes quietly.

the honest part

What we won't do

Some holes need a human decision — who should see this table, whether that endpoint should exist at all. When we hit one, we won't guess. You get a plain-language explanation of the hole, why it can't be fixed mechanically, and what the decision actually is. A tool that knows its limits is a tool you can trust with the rest. We'd rather tell you we can't than hand you a fix we can't prove.

05 / from builders

The opposite of the thing that broke your app.

06 / the front door

Point it at your live app. We'll show you what anyone can read.

No signup, no agent, nothing to install. Paste the URL of your deployed app and Meshbound runs the same first pass an attacker would — read-only, against your live endpoints, using only what your app already hands every visitor. If a table is open, we show you a redacted row out of it. Not a warning. The row.

Read-only. We never write to your app. Or see an example scan