Security
How your graph is protected, stated as what the software does rather than as a posture. Where a protection is missing it is named below rather than left out.
In transit
Every connection is TLS, and every hostname we serve declares strict transport security for itself and its subdomains, so a browser that has seen us once refuses to reach us any other way.
Sign-in
- Passwords are stored as a scrypt digest with a per-account salt and deliberately expensive parameters. We never hold the password, so we cannot tell you what it was.
- A wrong email and a wrong password give the same refusal, so the form cannot be used to discover which addresses have accounts.
- Sign-in attempts are rate limited per address and per account, counted against a one-way hash rather than the value, and the limiter refuses rather than waves attempts through if it cannot reach its store.
Tokens
- Access tokens are ES256, signed with a private key and verified against a published key set, so verifying a token never confers the ability to mint one. They last ten minutes.
- Each token names the surface it is for. A token issued for the application is refused at the connector endpoint and the reverse, so a token obtained for one cannot be replayed at the other.
- Refresh tokens rotate on every use. Presenting a spent one is treated as theft and revokes the whole family, ending every session descended from it.
- Signing out revokes immediately, on both the application and the connector, rather than waiting for the token to expire.
Connecting an assistant
The connector speaks OAuth 2.1 with PKCE and dynamic client registration, and the token it receives names the connector endpoint as its intended audience, so it is not accepted anywhere else.
The approval screen is served by the application you are already signed into, not by the api. There is no opaque request identifier for someone to send you, so you cannot be talked into approving a request you did not start. Every parameter the browser carried is revalidated on the server at the moment you approve.
Your own content
- A document body cannot be stored without being sanitized, because validation and sanitization are the same step at the write boundary. An assistant writing over the connector passes through the same check your browser does.
- Drawings are checked attribute by attribute against what each kind of shape is allowed to carry, and anything outside it is refused at the write rather than quietly dropped at the render. Scripts, stylesheets, embedded documents and external references are not permitted at all.
- Pasting a link fetches the page title through a guard that refuses private and internal addresses and rechecks every redirect hop, so a link cannot be used to make our server reach something it should not.
Permissions
Nothing you do not own is served to you, and the check is never folded into a database query: candidate rows are narrowed, the permission question is asked once, and only permitted rows are then loaded. Where a result is incomplete because permissions hid something, the response says how many were hidden rather than reporting a smaller total as if it were the truth, so a count cannot be used to probe for things you cannot see.
Every route is closed by default. Public ones are marked explicitly, so forgetting to protect a new route produces a refusal someone notices rather than an open endpoint nobody does.
At rest, and what that does not mean
The database, the cache and file storage are encrypted at rest by the providers named on the subprocessors page. That protects against a stolen disk. It does not make your content unreadable to us: titles and document bodies are stored as text, so an operator with database access could read them. Your graph is not encrypted end to end today, and whether to change that is an open design question rather than a thing we have quietly decided against. We would rather say so here than let the phrase encrypted at rest imply more than it means.
What we have not done
- End-to-end encryption of your content, as above.
- An external penetration test or a third-party audit. We have not had one.
- A compliance certification. We hold no SOC 2 report and no ISO 27001 certificate, and we will not imply otherwise.
Report a vulnerability
Email security@ideagraphs.ai, which is also published in our security.txt. Tell us what you found, how to reproduce it and what you think the impact is. We read every report and we will tell you what we decided and when a fix ships. We do not pay bounties.
In scope: the ideagraphs web application, the api and the connector endpoint, on hostnames we operate under ideagraphs.ai.
Out of scope: the providers on our subprocessors page and any system we do not run, since we cannot authorize testing against someone else's infrastructure. Also out of scope: denial of service and load testing, social engineering of us or of our users, physical access, high-volume scanning, and reports that are scanner output with no demonstrated impact.
Working with us. If you make a good-faith effort to follow this page, we will not start legal action against you over your research, and if someone else does we will say plainly that your work was authorized. To stay inside that: keep to the scope above, use only your own account and your own test data, stop as soon as you have shown the problem exists, do not access, keep or disclose anyone else's data, do not degrade the service for other people, and give us a fair chance to fix the issue before you discuss it publicly. This is our own commitment and not a waiver of anyone else's rights, and it cannot authorize anything unlawful.
Changes to this page
- 17 November 2026 First published.
Last updated: 17 November 2026