TRS.01
Security Architecture
The controls we implement in AI systems: isolation, permissions, secrets, logging and prompt-injection defence.
Open pageTrust & Security · 9 pages
Trust, evidenced.
Security review is where most AI projects stall. These pages say what we implement, what we can show, and what we do not hold — so a procurement team can check rather than trust a badge.
Introduction transcript. Client: How do I know I can trust the system? Qubi: Let me show you my working. Qubi: Documented controls. Traceable decisions. Evidence your team can check. Client: No shiny badge doing all the talking? Qubi: Just the things we can actually show you. Qubi: Bring your questions. I’ve saved you a seat.
At a glance
What we can and cannot show.
- 01 Certifications
- We do not currently hold SOC 2, ISO 27001 or HIPAA certification. Any certification we complete will appear on this site with its audit date and scope.
- 02 The controls
- We implement the controls those frameworks describe in every system we build, document them per engagement, and build systems that generate the evidence your auditors need.
- 03 Training on your data
- Where the provider offers it, we use enterprise endpoints configured for zero retention and no training on inputs — confirmed against the provider’s current terms per engagement.
- 04 Who holds the data
- Code and infrastructure live in your repositories and accounts. Every third party in the processing path is named, with its role and region.
WHAT WE CAN EVIDENCE TODAY
Certification status, plainly.
Certification status
We do not currently hold SOC 2, ISO 27001 or HIPAA certification.
Most agency sites in this space display those badges as design elements. We would rather tell you plainly, because a vendor claim you cannot verify is worth less than an honest one — and because procurement will check.
What we do is implement the controls those frameworks describe, document them per engagement, and build systems that generate the evidence your auditors need. Clients who hold certifications we do not can still use us: the systems we build are designed to sit inside their certified scope.
When we complete a certification, it will appear here with its audit date and scope, and in the machine-readable version of this site. Until then, this page says what is true.
THE LEDGER
What we hold, and what we do not.
Each row is a claim a procurement team can check, with the status it carries today.
-
SOC 2, ISO 27001 and HIPAA certification
We do not currently hold SOC 2, ISO 27001 or HIPAA certification. Any certification we complete will appear on this site with its audit date and scope.
Not held -
The controls we implement in every AI system we build.
Implemented -
Prompt injection treated as an authorisation problem
Retrieved content is designed to be treated as untrusted data rather than as instructions. Tools carry the minimum permission needed, destructive operations require confirmation, and retrieved content cannot authorise a subsequent tool call. Filtering helps at the margin but cannot be the primary control.
Implemented -
Controls documented for each engagement
We document them per engagement, and build systems that generate the evidence your auditors need.
Documented per engagement -
Zero retention and no training on inputs
Where the provider offers it, we use enterprise endpoints configured for zero retention and no training on inputs — at the time of writing, and confirmed against the provider’s current terms per engagement. Zero retention is at the provider; your own logs stay in your infrastructure under your retention policy.
Where the provider offers it -
Data residency set in discovery
EU regions for EU data residency, your own VPC where that is required, or entirely on your hardware with open-weight models where nothing may leave the perimeter. The requirement is captured in discovery because it constrains model selection.
Documented per engagement -
Code, infrastructure and fine-tuned weights are yours
Code in your repositories, infrastructure in your accounts, any fine-tuned weights yours. No proprietary platform and no runtime dependency on us.
Included in every engagement -
Sub-processors named
Every third party in the processing path, including model providers, disclosed with role and region. No unnamed “trusted partners”.
Included in every engagement -
Data processing agreements
Yes, and we can work from your paper. Standard contractual clauses where a transfer requires them. For engagements where we never hold your data, we say so rather than papering over a processing relationship that does not exist.
Available on request -
Security questionnaires
Where a control is not in place, the response says so. We would rather lose a deal on a truthful answer than win one on a carefully worded almost-yes.
Gaps stated, not hidden
Third-party terms are stated at the time of writing. We confirm them against the provider’s current terms for each engagement.
HOW WE WORK
Three areas, each with its own page.
-
01
Security architecture
Isolation, permissions, secrets handling, logging and prompt-injection defence — the controls we implement in every AI system we build.
-
02
GDPR & data residency
Lawful basis, residency, sub-processors and subject rights when a language model sits in the processing path.
Guidance, not legal advice GDPR & residency → -
03
The EU AI Act
Risk tiers, GPAI obligations and timelines — what deploying an AI system in the EU now actually requires.
Guidance, not legal advice The EU AI Act →
THE COMMITMENTS
What we commit to in writing.
01
We configure so your data is not training data
Where the provider offers it, we use enterprise model endpoints configured for zero retention and no training on inputs — at the time of writing; we confirm the provider’s current terms per engagement — and state the exact configuration in the architecture document. Zero retention is at the provider; your own logs stay in your infrastructure under your retention policy.
02
You own everything we build
Code in your repositories, infrastructure in your accounts, any fine-tuned weights yours. No proprietary platform and no runtime dependency on us.
03
Sub-processors are named
Every third party in the processing path, including model providers, disclosed with role and region. No unnamed “trusted partners”.
04
We answer questionnaires honestly
Where a control is not in place, the response says so. We would rather lose a deal on a truthful answer than win one on a carefully worded almost-yes.
- Building software since
- 2006
- Based in
- Barcelona
- Model stance
- Agnostic
- Code ownership
- Yours
Facts about how we work, not measured client outcomes.
Questions
What procurement asks us first.
Can you work with us if we require a SOC 2 certified vendor?
If it is a hard gate in your procurement policy, no — and we would tell you that on the first call rather than in week six. Some organisations can accept a risk exception for a small engineering supplier with documented controls; that is a decision for your risk function, not something we would try to argue you into.
Where is our data processed?
Wherever your requirements dictate. EU regions for EU data residency, your own VPC where that is required, or entirely on your hardware with open-weight models where nothing may leave the perimeter. The residency requirement is captured in discovery because it constrains model selection, and discovering it late causes rework.
How do you handle prompt injection?
As an authorisation problem, not a filtering one. Retrieved content — documents, tickets, web pages, emails — is designed to be treated as untrusted data rather than as instructions. Tools carry the minimum permission needed, destructive operations require confirmation, and the design does not allow retrieved content to authorise a subsequent tool call. Filtering helps at the margin but cannot be the primary control.
Do you sign DPAs?
Yes, and we can work from your paper. Standard contractual clauses where a transfer requires them. For engagements where we never hold your data — systems running entirely in your infrastructure — we will say so rather than papering over a processing relationship that does not exist.
TRUST — ALL PAGES
Every trust page.
-
-
TRS.02
GDPR & EU Data Residency
Lawful basis, data residency and subject rights when an LLM is in the processing path.
Open page -
TRS.03
The EU AI Act
Risk tiers, GPAI obligations and timelines — what deploying an AI system in the EU now requires.
Open page -
TRS.04
Data Residency & Sovereignty
Where data physically sits, who can compel access to it, and how each hosting choice changes the answer.
Open page -
TRS.05
Sub-processors
The vendors that may process personal data on our behalf, and why your system's providers are contracted by you.
Open page -
TRS.06
Model Supply Chain
Who trained a model, who operates it, what its licence permits, and how to evaluate a model you did not build.
Open page -
TRS.07
Data Processing Agreement
What our data processing agreement says in plain English, when it is needed, and what we will and will not sign.
Open page -
TRS.08
Security Questionnaire
Our standing answers to the questions every enterprise security review asks, published so you do not have to send the spreadsheet.
Open page
Qubi Your guideTrust guide
01 / 11
Shall I show you around?
Explore with Qubi
