Trust / AIpento

Security and retention

Last updated 6 October 2026. How the service itself is built, since we are asking you to trust us with read access to your code. Claims here are ones you could verify by asking.

Draft, not yet in force. Two details are still being confirmed before this page binds anyone: our registered address in Dubai and the mailboxes named below. Both are marked like this.

What the product does, and what it touches

AIpento does two things to one application. A code scan reads the repository you have given us read-only access to, and never runs any of it. A penetration test sends real requests to a staging address you own and have authorized in writing, and to nothing else. A finding reaches you only if it was reproduced against that running app; one that cannot be reproduced is withdrawn rather than reported with a lower score.

It is sold in two editions, which differ in the one respect a reviewer cares about - where the data sits, and what crosses your network boundary.

  • SaaS. We run it. Your repository is cloned onto our runner for the length of a run and discarded when the run ends. Findings, reports and audit records stay in our database.
  • Enclave. You run it, inside your own network, from a container image whose digest you pin. The database is a file on your own disk. We have no inbound path to it, nothing in it runs on a timer, and its licence is a signed file checked offline - there is no call home to validate it.

How your account is protected

  • Passwords are stored as PBKDF2-SHA256 derivations with 100,000 iterations and a per-user random salt. We never see, store or email a password, and we cannot recover one for you.
  • Session tokens are stored as SHA-256 hashes. A copy of our database does not let anyone replay a live session, because what is stored cannot be presented as a token.
  • Sessions last twelve hours. Deactivating someone revokes every session they hold immediately, not at the next expiry.
  • Invitations expire after seven days and password resets after one hour. Each is single use: consuming it destroys it.
  • Repeated failed sign-ins lock the account temporarily. The response to an unknown address and to a wrong password is identical, so the form cannot be used to discover who has an account.

How one customer is kept out of another's data

Every query that touches customer data takes an organisation identifier, and the helper that applies it refuses to run when that identifier is missing or empty rather than falling back to returning everything. An empty scope denies; it never widens. This is structural: there is no code path where forgetting the scope silently returns another customer's rows.

What runs your code, and when

  • A scan never executes anything. It clones the repository read-only, with any credential prompt disabled, and reads the files.
  • A verification run executes your test suite, and only that. It requires an explicit flag the operator must pass, a disposable container, and your written request. It is bounded in memory and in time.
  • Neither ever touches a production system of yours, and we never ask for production credentials.

Where your data lives, and what leaves

SaaS

Everything is held on Cloudflare: accounts, engagements, reports, findings, invoices, audit records and sessions in D1; report requests from the public form and rate-limit counters in KV. The table in the next section gives the retention for each. Two things cross that boundary during a run. The staging address you authorized receives the test requests - that is the test. And if you buy an AutoFix, the code being patched is sent to the model provider configured for your account, and the text of the finding is sent with it. Nothing else is sent anywhere.

Enclave

The appliance has a single outbound chokepoint, and every server-side request the image makes goes through it. It is off by default, and off means an outbound request is refused before a name is resolved or a socket is opened - including a request to the model vendor. An appliance that has not been configured at all refuses outbound traffic for the same reason. On a default install, nothing leaves.

Turning it on is a deliberate decision, because what it carries is your source code. When it is on:

  • The destination list is exact hostnames. There is no wildcard and no suffix match, so an entry for one vendor cannot be satisfied by a name that merely ends with it.
  • HTTPS only. A URL that carries credentials is refused outright rather than cleaned up.
  • Every redirect hop is re-checked against the same list, and authorization and cookie headers are dropped when a redirect crosses to a different host.
  • Once the name resolves, a private, loopback, link-local, multicast or cloud-metadata address is refused even when the hostname is on the list, and the connection is pinned to the address that was screened. The allow-list cannot become a route into your own network.
  • Every call writes one row to a transfer ledger: the host, the number of bytes, and which class of content it was - source code, finding text, or configuration. Never the content itself. A refusal writes a row as well, so what was blocked is as auditable as what was sent.

There is one other outbound path, and only if you configure it: audit events can be forwarded to a SIEM collector you name, so that the record lives somewhere a host administrator here does not control. It carries event metadata through a fixed list of fields - never a secret value, a prompt, or the body of a finding's evidence.

What is kept, and for how long

DataWhereRetention
Report requests from the public formCloudflare KV365 days, expired by the store itself rather than by a job we could forget to run
Accounts, engagements, reports, findings, invoicesCloudflare D1Life of the account; invoices longer where tax law requires
Audit recordsCloudflare D1, append onlyLife of the account
Session recordsCloudflare D1, hashed12 hours
Your repository checkoutTemporary disk on the runnerDiscarded when the run ends
Rate-limit countersCloudflare KV, salted hash of the IP1 hour

Backups

Backups are taken as a structured export of the database. Two things are deliberately left out of every backup: password derivations are redacted, and live sessions are skipped entirely. A restored copy therefore cannot be used to sign in as anyone - everyone sets a password again. A restore refuses a backup whose row counts do not match its own manifest rather than applying half of it.

How the site is served

Everything is served over HTTPS with HSTS. The content security policy allows scripts only from this origin and from Cloudflare Turnstile, and forbids framing entirely. No page assigns HTML from a value that came from a server response; every element is constructed as a node, so a value can never become markup.

What we do not claim

  • We hold no certification. There is no SOC 2 report and no ISO 27001 certificate, and we will not imply otherwise to win work.
  • The static report proves the presence of the patterns it finds, never the absence of others.
  • Dependency checks cover npm and PyPI packages, looked up in the public OSV vulnerability database. Packages from other ecosystems are not checked yet.

Telling us about a problem

If you find a weakness in this service, write to [email protected]. We will confirm receipt within two working days and tell you what we intend to do. We will not threaten you, and we will credit you if you want the credit.