Security
We deploy inside your cloud. Your data never leaves your environment.
Systems run in accounts you control, under model API terms that exclude training on your data. This page states exactly what we do, what we align with, and what we do not claim.
Last updated 2026-06-10 · Security contact: James Booth · info@advizr.ca
One client, one database. No shared anything.
Authentication is isolated from client data.
Login and identity run in their own database that holds no client business data. A compromise of one layer does not expose the other.
Each client gets an isolated database.
Your engagement runs on its own database instance. No shared tables, no pooled storage, no schema where a bad query could cross tenants.
There is no cross-client access path.
Separate instances, separate credentials, separate knowledge stores. A query against one client's systems cannot reach another's, by construction rather than by policy.
What we commit to, on every engagement
01
Your data is never used to train models.
We use commercial API tiers from Anthropic and OpenAI whose terms exclude training on API inputs and outputs by default. Both publish this on their own trust pages, so you can verify the chain instead of taking our word.
02
Zero data retention where required.
For engagements that need it, we negotiate zero-data-retention terms with the model providers so prompts and outputs are not retained on their side.
03
Data minimization before any model call.
Personal information is redacted or minimized before anything reaches a model API. The model sees what it needs to do the job, not your whole database.
04
Least-privilege keys, scoped per environment.
Every integration gets its own credential with the narrowest scope that works, revocable per client. No shared master keys across clients or environments.
05
Audit logging you can read.
AI interactions and system actions land in per-client audit tables inside your own database. The trail lives in your accounts, not ours.
06
Human approval gates on high-risk actions.
Irreversible or externally visible actions route to a person before they execute. Agents assemble, retrieve, draft and monitor; people approve.
07
Breach notification, stated up front.
If an incident touches your data, we notify you promptly with what we know and what we are doing about it, so you can meet your own regulatory clocks, including Quebec Law 25 timelines where they apply to you.
08
Data processing agreement on request.
We sign DPAs. Email the security contact below and you will get one back, along with our security one-pager.
Canadian data residency, named
Where your data must stay in Canada, these are the concrete, verifiable options we deliver on.
Your own cloud (default)
The default posture: systems deploy into accounts you control, so residency follows your cloud. We hold operator access you can revoke, not custody of your data.
Azure OpenAI, Canada East and Canada Central
Inference and data processing inside Canadian data centers, with model-availability caveats that we scope per engagement.
AWS Bedrock, ca-central-1
Data at rest stays in Canada. Where cross-region inference applies, traffic travels the AWS backbone, never the public internet.
Built against the OWASP Top 10 for LLM Applications (2025)
All ten risks, mapped by name to what we actually do about them. The list itself →
LLM01
Prompt injection
Defense in depth, because no single fix exists: input validation, context isolation, system-prompt hardening, output filtering, least-privilege tool access, and adversarial testing before launch.
LLM02
Sensitive information disclosure
PII minimization before model calls, retrieval scoped to what the requesting user may see, and no secrets in any prompt.
LLM03
Supply chain
Pinned dependencies, vetted model providers under commercial terms, and a published subprocessor table so you can audit the chain.
LLM04
Data and model poisoning
Retrieval corpora are curated from your own sources with provenance tracked. We do not fine-tune on unvetted data.
LLM05
Improper output handling
Model output is treated as untrusted input: validated and escaped before it touches downstream systems, never executed directly.
LLM06
Excessive agency
Agents get the narrowest tool set that does the job, and human approval gates sit in front of irreversible actions.
LLM07
System prompt leakage
Prompts are assumed readable by a determined user. Credentials and sensitive logic live in environment configuration and code, never in the prompt.
LLM08
Vector and embedding weaknesses
One isolated vector store per client, access-controlled at the database layer. No shared embedding space across clients.
LLM09
Misinformation
Generation is grounded in retrieved sources with citations, checked by evals against golden datasets, and human-reviewed where the output matters.
LLM10
Unbounded consumption
Rate limits, per-key quotas and budget alerts on every model integration, so a runaway loop is a paged alert, not a surprise invoice.
Framework alignment
We claim no certifications we do not hold.
Alignment is a practice claim and we make it carefully. Certification is an audit outcome and we do not assert ones we have not earned.
NIST AI RMF
AI risk management mapped to the four functions: Govern, Map, Measure, Manage. Voluntary by design, and it maps directly onto how we scope and run engagements.
SOC 2 Trust Services Criteria
Controls aligned with SOC 2 Trust Services Criteria: security, availability, confidentiality. Mapped policies, not an attestation.
ISO/IEC 42001
ISO/IEC 42001-informed governance for how AI systems are specified, monitored and retired. Both of our model providers hold the certification itself.
Built to respect the law you answer to
PIPEDA
Delivery designed to support your obligations under PIPEDA: accountability and individual access for AI processing of personal information, with minimization built into every pipeline.
Quebec Law 25
Transparency for automated decisions and tight breach-notification clocks. Our incident posture and audit logging are built so you can meet them.
BC PIPA
We are a BC company. Provincial private-sector privacy law is our home regime, and the one our own operations answer to.
EU AI Act
Scoped per engagement: we determine whether a build makes you a deployer or a provider, and what Articles 4 and 50 require of you, before we write code.
Subprocessors
The chain, published
Buyers verify chains, not vendors. Every posture below restates what that vendor publishes on its own trust page, last checked 2026-06-10. Facts, not partner logos.
| Vendor | Role | Published posture | Verify |
|---|---|---|---|
| Anthropic | Model API (Claude) | No training on API data by default; ZDR available. SOC 2 Type II, ISO 27001, ISO/IEC 42001. | Trust page → |
| OpenAI | Model API (GPT) | No training on API data by default; ZDR on eligible endpoints. SOC 2 Type 2 report available. | Trust page → |
| Supabase | Database and auth | SOC 2 Type 2; HIPAA configuration available. | Trust page → |
| Vercel | Web hosting | SOC 2 Type 2; ISO 27001. | Trust page → |
| Modal | Serverless compute | SOC 2 Type 2; external penetration testing. | Trust page → |
The questions procurement asks
Need the paperwork? Ask the person who signs it.
DPA, security one-pager, or your own questionnaire. James Booth answers security email himself.