Trust · Security model
Secure by code, not by prompt
Brigham.ai is an AI that does real work in your systems. So the rules that keep it safe are written in code the model can’t edit, talk around or be tricked out of.
The model proposes. Code decides.
A prompt is a request. Code is a rule. Brigham.ai treats every instruction to the model as advice that can be ignored or overridden, and puts anything that matters in a layer the model never touches.
Eight controls, each enforced in code
Each one has its own page with the mechanism, not just the promise.
Policy is code
Every tool call runs one pipeline: look up, validate, check policy, approve if needed, execute, validate, redact, audit. There is no other path.
Action classes
Read, reversible write and sensitive. Sensitive actions always wait for a person.
Untrusted content
Email, web pages, documents and notes are data. Nothing that arrives through a tool can approve an action or widen a scope.
Credentials
Encrypted at rest and referenced by id. They never enter the model’s context, the logs or a tool result.
Your inbox is yours
Each person’s mail and calendar are private to them. Workspace admins have no access.
Protected domains
Finance access comes from a grant to a named person, never from a role. The label follows the data.
Tenant isolation
A workspace is a company, and a company is a security boundary. Every row carries the workspace id and every query is scoped by it.
Export and deletion
Self-serve tools to export and delete your data, with every request audited.
Some data is never ingested
Some categories of sensitive data are excluded from Brigham.ai altogether. Health information is one example. The rule is simple: if the data is too sensitive to risk, Brigham.ai does not take it in, so it cannot leak, be summarized into the wrong place or be sent to a model. The Terms of Service prohibit loading excluded categories unless we have agreed otherwise in writing.
- Excluded at the source, not filtered later
- Applies to every workspace
- Stated in the Terms of Service
Two more guarantees
Delivered means delivered
A mail provider accepting a message is not delivery. Brigham.ai reports “sent” when the delivery event arrives, not when the provider returns a 202.
Budgets are checked before the call
Per-run token and time caps and a per-workspace daily spend cap are enforced in code before each model call. A runaway loop stops itself.
What we claim today
We would rather be exact than impressive.
- Certifications
- None. We do not claim SOC 2, ISO 27001 or any other attestation.
- Evidence
- We describe what the system does rather than pointing to badges.
- Hosting
- Cloud hosting in the EU.
- Sign-in
- Through erp.io. Brigham.ai has no passwords of its own.
Read the mechanism
Questions, answered.
Is Brigham.ai certified?
No. We don’t claim SOC 2, ISO 27001, HIPAA or similar. We describe what the system does and let you check it.
Can I ask for more detail?
Yes. If you’re reviewing Brigham.ai for your company, request access and ask for the security questionnaire conversation. We would rather answer specifics than send a badge.
Review it with us
Request access and bring your security team. We’ll walk through any control on this page.