Security
What actually protects your account and your thinking. Everything below describes measures in the running code — where something is not yet done, this page says so.
Last reviewed August 2026 · Report a vulnerability: hellosocria@gmail.com
Your account
Sign-in is handled by Clerk rather than by us. We never see or store your password. Session handling, multi-factor authentication where you enable it, and device management are theirs — a specialist doing one job well beats us reimplementing it.
Payments run through Stripe’s own hosted checkout. Card numbers never touch our servers; we hold only a customer reference and your subscription status.
Where your data sits
Conversations, maps, drafts and memory live in a managed Postgres database at Supabase, encrypted at rest and in transit by the host. The browser never talks to that database directly. Every read and write goes through our server, which checks on every request that the row belongs to the account asking for it. That is not a rule applied to deletes and hoped for elsewhere: every query in the codebase is filtered by account, updates included, and the routes that call the AI hold no database state at all, so there is no record identifier a request could substitute to reach someone else’s thinking.
The database has row-level security switched on with no policies and the public API roles revoked, which is a second wall rather than the first one. If a key ever leaked, or the database’s own REST interface were exposed by mistake, the answer would be nothing rather than everything.
Your messages are encrypted in transit and encrypted at rest by the database host, and access to them is controlled per account. They are not end-to-end encrypted. We could read them if we chose to, and so could our AI provider in the course of generating a reply. Anyone telling you otherwise about a product that builds a Thinking Map from your words is selling something.
If you ever connect Google or Notion, those access tokens are encrypted before they are stored — AES-256-GCM, with a key derived from a server-side secret. Without that secret the system refuses to store a connection at all rather than fall back to plaintext — a database dump on its own does not yield anyone’s Google or Notion access. Finishing a connection also checks a signed, single-use link against the session making the request, so a link cannot be re-pointed at somebody else’s account. Disconnecting deletes the stored credential.
What the server refuses to trust
Most of Socria’s security work is in what it declines to take at face value.
- The browser is never the authority on what you have paid for. Every gated request re-decides your plan on the server, from the payment processor and your account — the client only states what it believes.
- Payment webhooks are verified before they are read. The raw body is checked against a signing secret before anything is parsed. A forged event, a replayed old one, or a genuine signature reused over an edited body are all rejected.
- The price comes from the server, never the request. A client that could name its own price could name its own number.
- Model output is treated as untrusted. Everything the AI returns is validated and bounded before it reaches your screen. When Logos shows you a quotation of your own words, the model supplies only the position in the transcript — the text is taken from your real transcript by the server, which is why a fabricated quote is not merely unlikely but impossible.
Handling risky input
- Maths expressions are parsed, not evaluated. Plotting a function means running a string you typed. Socria uses a hand-written parser over a fixed grammar instead of the language’s own evaluator: a hostile expression can produce a meaningless number, it cannot run code.
- Fetching a web page is screened. When you point Socria at a URL, the request is blocked from reaching internal addresses — loopback, private ranges, link-local and cloud metadata endpoints — including tricks like integer-encoded IPs, and the check is repeated after redirects.
- Rendered maths cannot inject markup. The formula renderer runs with untrusted input in mind and hard limits on how far an expression can expand.
- Rate limits apply per account, or per IP when signed out, across separate budgets for expensive and cheap operations.
What we have not done yet
Socria is a young product. We have not completed a third-party security audit or a penetration test, we do not hold SOC 2 or ISO 27001, we do not yet run a formal bug-bounty programme, and we have not signed a data processing agreement with our AI provider or moved to a zero-retention endpoint. Your messages are not end-to-end encrypted. If your organisation needs any of those before adopting a tool, we are not there yet — and would rather tell you than let you assume.
Reporting a vulnerability
If you find something, please tell us at hellosocria@gmail.com before disclosing it publicly, and give us a reasonable window to fix it. We will confirm we received your report, keep you posted, and credit you if you would like that. We will not pursue legal action against anyone acting in good faith under this policy.
For what we collect and how to have it deleted, see the privacy policy.