Security

Security: how access, data and automation are controlled

Updated Markdown
In short

Cureman signs people in with a one-time code and has no passwords. Every page and action checks a permission, every query is scoped to the organisation, and secrets such as connector credentials are stored encrypted. Connectors only read, every change is written to an audit log, and what a model reads is treated as data. An agent never sees a login, and an assistant key can read and propose only. This page describes what the code does and makes no certification claims; questions about hosting and contracts go to the operator.

This page lists the security controls built into Cureman, as the code implements them. It makes no certification statements. The platform is deployed by its operator, so questions about hosting, contracts and data processing go to the operator.

Sign-in by one-time code

People sign in with a one-time code sent to their email address. There are no passwords. Registration is closed: a code is sent only to an address that already belongs to an organisation, and the form answers the same way whether or not it knows the address. Sessions are signed cookies that scripts in the browser cannot read.

Roles and permissions

Every page and every action checks a permission, never a role name. The built-in roles follow jobs: owner, admin, procurement, finance, IT admin, budget owner, auditor and member. An organisation can define roles of its own from the same permissions, and nobody can create or grant a role that holds a permission they do not hold themselves.

Prices, costs and savings are shown only to people with the permission to see figures. An organisation can also require a second person for everything that leaves it, so whoever drafted a message cannot approve it.

Tenant isolation

Every query on an organisation's data is scoped to that organisation, which is taken from the signed-in session, including lookups by id. An id from a form is never trusted on its own. Pages and actions of a module are additionally refused when the organisation has not booked that module.

Secrets

WhatHow it is kept
Connector credentials, keys for model providers, Slack secrets, mailbox tokens, logins in the access vaultEncrypted with AES-256-GCM
Gateway keys, assistant keys, runner keys, ingest keysOnly a hash is stored

Secrets are not sent to the browser and not placed in an address.

Connectors only read

Connectors can reach only the hosts their definition declares, over HTTPS. A change at a vendor is a step of an action plan: prepared, shown in full, approved by a person, and off unless the organisation allows it.

The audit log

Every change is written to an audit log with who made it and what changed. What the platform did by itself is recorded there too, marked as automatic. Support access by the operator's administrators appears in the organisation's log, marked as such.

An organisation's own host names

An organisation can use host names of its own for the dashboard and the gateway endpoint. A host name is trusted only after a DNS record has proved control of it. Requests arriving on it are tied to that organisation: someone who is not a member is refused, and the gateway accepts only that organisation's keys.

What a model reads is data

An invoice, a supplier's reply, an employee's answer or a web page is passed to a model as data, with the instruction to treat it as such. The output must fit a fixed shape, is validated before use, and is displayed as plain text. Reading a PDF or a reply means its text goes to the model provider the operator has configured; structured e-invoices and CSV files are read without a model.

The gateway stores token counts, the task type and a user reference for each request. It stores no prompt or response text unless the organisation switches samples on. From a connected mailbox only replies to conversations started in the platform are stored. The access a mail provider grants still covers the whole mailbox, so a mailbox made for purchasing is a better choice than a personal one.

The access vault and the agent runner

Logins for websites of vendors and suppliers are kept in the access vault. A login is released to the runner only for a task that is running, only when the entry allows agent use and its hosts cover the hosts of the task. The runner types it into the page; the model never sees it. The agent enters exactly the values a person approved, stays on the hosts of its task and by default stops before a submit until a person confirms. Every step is recorded, and agent tasks are off unless the organisation allows them.

Assistant keys

An organisation's own assistant connects over MCP with a key that can be revoked. A key reads purchasing data, optionally without amounts, and with the scope to propose it can leave a proposal that a person decides. No tool sends, approves, awards or changes anything, and each stored proposal is audited under the key's name.

Demo organisations

Organisations built for sales calls are fictional: their people and suppliers have addresses at .example, and they never send mail. The operator keys that build them reach demo organisations only.

The rules that govern what the AI may do are listed in the principles.

Questions and answers

How do people sign in?

With a one-time code sent to their email address. There are no passwords to store, reuse or reset. Registration is closed, so a code is only sent to an address that already belongs to an organisation, and the sign-in form answers the same way whether or not it knows the address.

Can a connector change something at a vendor?

No. Connectors only read, and they can only reach the hosts their definition declares. A change at a vendor is a separate step of an action plan that is prepared, shown in full and approved by a person, and it is off unless the organisation allows it.

Does the AI agent see our logins?

No. Logins are kept encrypted in the access vault and are released only to the runner, only for a running task whose hosts the login covers. The runner types the login into the page itself, the model that drives the browser never receives it, and every release is written to the audit log.

What can an AI assistant connected over MCP reach?

Purchasing data of its organisation, and with the matching scope it can leave a proposal that a person decides. A key can be made blind to amounts. No tool offered to an assistant sends, approves, awards or changes anything.

Where is Cureman hosted, and who signs the contracts?

Cureman is deployed by its operator. Where it runs, how backups are kept and what a data processing agreement says are questions for the operator, and this page makes no statement about them.