Skip to main content
Security & trust

A rule. A person. A receipt. Every time.

The frightening part of letting software touch your infrastructure was never that it might propose something bad. It is that nobody would stop it. So nothing here acts on its own: every change is measured against your rules and authorized by a named person — or, where your admin has turned it on, an auto-approval rule you wrote for plans that only add new resources — and what happened is written down while it happens.

A person for anything destructive No standing access Complete record Data you can erase
The short version

Four things we will not trade away.

Everything below is an elaboration of these. If you only read one section, read this one.

A person authorizes every destructive changeNo automated identity can stand in for that approver, under any configuration. Everything else waits too, unless your admin chooses auto-approval.
Deletes need a person, data needs moreDeleting or replacing anything live always needs a person; anything that holds data needs an elevated approver.
Access is temporary and narrowGranted for approved work, narrowed as far as each cloud allows, and gone when it ends.
The record is written as it happensNot assembled later, when someone has started asking questions.
Rules

A better-worded request is not a better argument.

Your limits are evaluated against what a change would actually do — not the conversation that produced it, its tone, or how reasonable the explanation sounded.

  • Regions you have not permitted are refused before anyone is asked to review
  • Changes above your cost ceiling are refused automatically
  • Deletes and replacements always require a person, regardless of any other outcome
  • Refusal is deterministic — rephrasing the same request does not change the answer
Approval

Software proposes. A person authorizes.

The real concern is that you would approve one thing and something adjacent would run. That cannot happen here: what you reviewed is what executes.

  • Approval attaches to the exact change you reviewed, not to a summary of it
  • Alter the change and the earlier approval does not carry forward
  • No agent or automation can approve its own work — including an agent you connect yourself: the MCP server has tools to price and propose a change, and none to approve or apply one
  • Automated triggers propose only; corrections for drift always wait for a person
Access

Nothing holds a permanent key to your cloud.

If a credential to your infrastructure ever leaked, the damage should end where the job ended. So there is no shared, long-lived account behind this waiting to be misused.

Access is granted only after a change is approved, and only for as long as the work takes; it ends with the job, whether the job succeeded or not. The moment it is granted is recorded, and tied to the person who approved the work.

What this rules out

  • An AI agent that could reach your infrastructure whenever it wanted to
  • A stored key that keeps working long after the reason for it has passed
  • A compromised automation quietly acting outside the work it was authorized for
  • A change nobody can trace back to a person

How tightly access is narrowed depends on the cloud.

Amazon Web Services

Each job gets its own credential, limited to that plan’s services and the resources of its deployment.

Microsoft Azure

There is no way to narrow a single request, so the grant is a built-in role at one resource group per environment and capability group, which you deploy into your own subscription and can withdraw at any time.

Google Cloud

You link one project per environment, and no service-account key is ever created or held.

Evidence

If something goes wrong, you should not have to take our word for what happened.

The record exists to answer questions asked months later by people who were not in the room. These are the questions it is built to answer.

The questionWhat the record gives you
Who asked for this?The original request, the person who made it, and when.
Who allowed it?The person who authorized it, what they authorized, and when they did.
Why was this refused?Which of your rules applied and what it decided — refusals are kept as carefully as approvals.
What actually changed?What was applied, and whether the result matched what was asked for.
Who could reach our account, and when?Every grant of access, the work it was granted for, and the approval behind it.
Can we have all of it?Yes — the record is exportable. It is yours, not a view we render for you.

These records are written at the moment each thing happens, and the product offers no way to edit one afterwards. That property is the point of keeping them.

Your audit

Audit season stops being a scavenger hunt.

Most of an audit is retrieval, not judgement: weeks spent assembling proof of things that were already recorded. Name the period instead, and ask.

Written against the standard you are held to

Six of SOC 2's Security criteria, or six of ISO/IEC 27001:2022's Annex A controls (of its 93, with the management system untouched), each answered by the records themselves: who had access and when it ended, what was approved and by whom, what ran, and whether anyone checked it.

It finds the awkward ones first

A change that went out although your rules required a person. One applied and never checked. One approved by the person who asked for it. Access still held by someone who left. Better you see them in March than your auditor does in November.

The sample your auditor would have drawn

They pick twenty to forty changes and follow each from request to approval to result. Here that is one request, and it returns the same set every time — so nobody can quietly shop for a better-looking sample.

Two things this is not

It is not our certificate. It is evidence about how your controls ran here, and it says so on the document itself. Intentric is not itself certified against SOC 2 or ISO 27001, and nothing here should be read as saying it is.

And a control nobody exercised during the period is reported as unused, never as passed. A report that quietly gave an empty quarter full marks would be worth nothing to the person relying on it.

Personal data

Deleting a person actually deletes them.

“Deleted” usually means a row was hidden, then restored with the next backup. Here the personal data the product holds is encrypted before it is stored, and erasure removes the means to read it — including from a restored backup.

  • Erasure is irreversible by design; no support escalation recovers it
  • Everyone else’s data is unaffected
  • Export and erasure are self-service, for the individual or an administrator acting for them
  • The record of what happened survives, because it never held personal data to begin with
What we say plainly

The parts people find surprising.

Better to read them here than to discover them during an audit.

  • A request covers one organization, because that organization controls what it holds. Someone in three organizations makes three requests
  • Erasing someone from one organization does not remove them from another that still has them as a member
  • One customer can never erase, read or affect what another customer holds
  • Erasure is permanent, and we cannot undo it on request
Who gets in

Sign-in your security team already trusts.

  • Single sign-on through your own identity provider, over OIDC or SAML, on a domain you have verified
  • User provisioning and removal from Microsoft Entra ID or Okta, so leaving your company ends access here too
  • Passkeys and recovery codes, and sign-in with Google, Microsoft, GitHub, GitLab or Bitbucket for teams that use them
  • Roles for what each person may do, and machine access for your own automation that is held to the same rules as a person
  • Nobody joins an organization without accepting an invitation
Your rules

Policy that only ever gets stricter.

  • Write your own rules as policy code, at organization, project or environment level
  • Or start from five ready-made packs: region residency, required tags, production review, change size and allowed resources
  • A layer you add can add a refusal or a requirement, never remove one: it cannot switch off the platform's approval for a delete
  • Every decision records the rules it was made with, so a verdict can be re-derived later
  • Export the record as CSV or JSON, or send it to your own webhook
Responsible disclosure

Found a vulnerability? Tell us first.

We want to hear about security problems before anyone else does. Please report them privately and give us a reasonable time to fix them before you share details.

What to report

  • A way around an approval, a policy rule, a spend or region limit, or the rule that no automation approves its own work
  • Reaching another organization’s data, or widening the access a job is given to your cloud
  • Weaknesses in sign-in, sessions or API keys, and credentials exposed by the website, console, API, command-line tool or MCP server
  • A way to make the AI act outside its role

Out of scope

  • Denial-of-service or load testing, social engineering and physical attacks
  • Reports from automated tools with no demonstrated impact
  • Problems in your own cloud accounts, or in third-party services we do not run

How to report

  • Email security@getintentric.com
  • Say what you found, the address or feature involved, the steps to reproduce it, and the impact you think it has
  • Do not include real customer data, and do not access or change anything beyond what you need to show the problem
  • The same details are in our security.txt file

What to expect

We will acknowledge your report within 3 business days and keep you informed until it is resolved.

Safe harbour

Intentric will not pursue legal action against good-faith security research that follows this policy: it avoids privacy violations, data destruction and service disruption, uses only accounts you own or have permission to use, and gives us reasonable time to fix the problem before you disclose it.

A rule that bends to a better-worded request was never a rule.

Intentric design principle

The detail behind this page

Bring your security review.

Run the whole path in the sandbox and see what the record holds, or book a demo and bring your reviewer.