Security and data

How Vythos is secured

Your rules are among the more sensitive things an agency owns: what you charge, what you've committed to, what you've agreed with each client. Here's how that material is kept.

Workspaces are isolated in the database

Every row Vythos stores carries the workspace it belongs to, and that isolation is enforced by the database itself rather than by application code remembering to filter. The account the application connects with cannot bypass it.

The practical meaning: a bug in a query can't return another workspace's material, because the database will not serve it.

Inside your workspace

• Rules are shared, filtered by access. Each person has a level for each category, and a category set to no access is never sent to them — not on screen, not to their AI tools. See Your team and their access.
• Memory is private to each person. Your uploads and your connected accounts feed your memory only. Nobody else can search it, the workspace owner included.
• Connections are personal. Each person authorises their own Google, Notion or GitHub account, and Vythos reads with exactly that person's access.

Connector access is encrypted at rest

The tokens that let Vythos read your Drive or Notion are encrypted before they are stored, with authenticated encryption, and are decrypted only to make a request on your behalf. They are never written to logs, never returned by the API, and never shown in the app.

Sign-in

• Passwords are hashed with bcrypt. Nobody at Vythos can read yours.
• Repeated failed sign-ins for an address are throttled, and the throttle is checked before the password, so a refused attempt reveals nothing about whether an account exists.
• Password reset is by emailed link.

What reaches a model

Vythos uses large language models for two things: writing an answer in Ask, and reading values out of a document when you add a rule. A separate model turns the text of each document into a search index (embeddings) when it's read in, and each question into the same form when you search.

• The models run on Microsoft Azure (Azure OpenAI), in the EU: requests are processed only in EU member states.
• What's sent is the material needed for that request — for a question, the rules and passages relevant to it, your agency's description and client list, and the last few turns of your own conversation; the text of the document you're adding; or, for indexing, the passages of a document.
• Microsoft's terms are that prompts and answers are not used to train models, are not available to OpenAI or any other model provider, and are not available to other customers. Vythos does not train anything on your material either.
• Azure runs automated abuse monitoring. Content it flags may be kept for human review, stored in the EU and reviewed only by authorised Microsoft staff in the European Economic Area.
• The rest of the product — following documents, versioning, permissions, checking a draft against values — involves no model at all. Draft checks in particular are deterministic: the same draft and the same rules produce the same result every time.

What Vythos stores, and what it doesn't

Stored: the text of documents you connect or upload, split into passages and indexed; your rules and their history; your clients and your agency's description; your team and their access; each person's Ask conversations, readable only by them; records of what was handed to which AI tool, so a draft can be checked against it.

Not stored: your files themselves. Vythos reads documents where they live and keeps the text; it never becomes the place your documents are kept.

Outside services involved

Service	What for	What it sees
Microsoft Azure — Azure OpenAI, EU	Generating answers, reading values from documents, indexing text for search	The text needed for that request; each passage of a document when it's indexed
Google, Notion, GitHub	Reading the documents you connect	Only what the account you connected can see, and only what you selected
An email provider	Invitations and password-reset links	The address and the message

Reporting something

If you think you've found a security problem, email us rather than opening a public issue, and give us a way to reproduce it. We'll confirm we've received it and tell you what we're doing about it.

Next: Your data.