Security and trust
Your proposal stays with the people you add to it.
A consortium proposal holds your partners’ plans, budgets and strategy. Cortyne is built so that access is denied unless it is explicitly granted, so that AI works on your terms, and so that a mistake in one place cannot expose another.
Private by default
Row-level security on every business table. Nothing is readable unless a rule allows it.
Partners see only their part
Each partner sees their own tasks and what you share. Not each other’s submissions.
No training on your data
Cortyne AI runs on Microsoft Azure OpenAI. Prompts and completions are not used to train models.
AI that asks first
It drafts and summarises. Nothing goes to a partner without your approval.
Access control
Denied unless explicitly granted.
Access is enforced in the database itself, not only in the screens, so it holds even if an interface has a bug.
- Row-level security on every table. Cortyne is built on Supabase (PostgreSQL). Every business table has row-level security switched on, so a query only returns the rows the signed-in person is allowed to see.
- Deny-by-default permissions. New tables and functions start with no access for anonymous or signed-in users. Each permission is granted deliberately, and policies are a separate control from those grants.
- Proposal access is explicit. You choose who is an owner, who has full access and who is only a partner. Being in the same organisation, or being an organisation admin, does not open a proposal.
- No master key for ordinary requests. Everyday requests run as the signed-in person, under their permissions. Secrets stay on the server and never reach the browser.
- Separate staff authority. Cortyne Team access is provisioned deliberately, is never derived from a profile or organisation membership, and does not grant access to customer proposals. Publishing funding-call information also requires multi-factor authentication.
- Accounts are reviewed. New organisations are checked by the Cortyne Team before their first proposal. Invited partners reach their brief through a secure link.
Cortyne AI
Microsoft Azure OpenAI. No training on your data.
Cortyne AI summarises partner EOIs, compares responses, answers questions about a proposal and drafts reminders. This is how it is set up.
- Azure OpenAI, not a public chatbot. Requests go from Cortyne’s servers to a model deployment in Cortyne’s own Microsoft Azure environment. Customer data is not sent to a consumer AI service.
- No stored API key. Cortyne authenticates to Azure with Microsoft Entra ID. There is no long-lived key in the code or the browser to leak.
- No training on your data. Microsoft’s published terms for Azure OpenAI say that prompts, completions and embeddings are not used to train foundation models, are not available to other customers and are not available to OpenAI. Cortyne does not train models on customer data either. Details are in Microsoft’s data privacy documentation.
- No provider-side conversation storage. Cortyne asks the provider not to store responses (
store: false). Cortyne keeps only the structured result and the audit details needed to trace and reuse it. - Only the data a task needs. Each AI task is given a deliberately limited context. For an EOI summary, contact names and emails, internal notes, other partners’ responses and unrelated budgets are left out.
- It sees only what you can see. AI tasks run as the signed-in person, under the same row-level security as the rest of Cortyne. Ask Cortyne on a proposal requires owner or full-access rights.
- Checked and traceable. Answers are validated against rules, for example against claiming that an action was taken. Every run is recorded with who ran it, the model, the prompt version, token counts and a fingerprint of its input. Nothing is sent to a partner without your approval.
One thing to know. Like other Azure OpenAI customers, Cortyne is subject to Microsoft’s abuse monitoring, under which Microsoft may store prompts and completions in a logically separated store for review by authorised Microsoft staff, as described in the documentation linked above. That data is not used for training.
Preventing data leakage
Several independent layers, so one failure is not a breach.
Each layer assumes the one before it could fail.
- Sign-inSecure cookies and server-checked sessions. Client claims about role or ownership are never trusted.
- Server rulesEvery read and change goes through server-side services that validate input and lifecycle steps.
- Database rulesRow-level security decides which rows each person can read or change.
- Minimal sharingPartners get only their own tasks. A lead organisation’s name can be withheld from partners when the manager chooses.
- RecordsImportant actions leave an audit trail. Published funding-call versions cannot be altered.
- Files are private. Uploads are size- and type-checked, stored privately and downloaded only through authorised links.
- Imports are untrusted. Structured imports are validated and cannot make themselves official. A person must verify and publish them.
- Atomic changes. Related updates and their audit records succeed or fail together, and Cortyne guards against duplicate and out-of-date edits.
- Tested on every change. Automated tests, including tests that check people cannot read each other’s data, run before each release. This is useful assurance, not independent certification, and we do not present it as one.
Straight answers
Questions your security team will ask.
Does Cortyne train AI models on our proposals?
No. Cortyne does not train models on customer data. Cortyne AI uses Microsoft Azure OpenAI, and Microsoft states that prompts and completions are not used to train foundation models, are not available to other customers, and are not available to OpenAI.
Can other organisations or partners see our proposal?
Only the people you add to a proposal can open it. Membership of the same organisation does not give access. Partners see their own tasks and what you choose to share, not each other’s private submissions.
Can Cortyne staff read our proposals?
Cortyne Team authority is a separate, deliberately provisioned context. It exists to publish and verify funding-call information and does not grant access to customer proposals.
Does Cortyne AI act on its own?
No. Cortyne AI summarises, compares and drafts. Nothing is sent to a partner without your approval, and answers come only from the proposal’s own records.
Is Cortyne independently certified?
Not yet, and we do not claim it. Our controls are tested automatically on every change, which is useful but is not independent certification.