Security
Security you can check, not just read.
This page is for the people who approve suppliers: information security, model risk and procurement. Each statement on it is tied to the code, configuration or audit evidence that makes it true, and an engineer will show you that evidence.
Your data kept apart, and people in control.
- Your data kept apartEach customer has its own Azure SQL database. Which database a request reaches is decided on the server from the signed-in user, never from the request, and cross-tenant tests run on every change.
- No email sent by AINo AI agent can send email. Replies are saved to your Drafts folder and you send them yourself. A build check fails if a send path appears in agent code.
- Agents get only the tools they are givenEvery tool is classified by what it can change. An agent that reads outside content, such as email, web pages or uploaded documents, cannot be given a tool that writes or deletes, and this is checked before any change is deployed.
- Approval by the named personA step that needs approval can be approved only by the person it is assigned to. Anyone else is refused, the work does not move, and the refusal is logged.
- Automation limited to filingRules your team sets up can move, flag, mark read and archive mail. Forwarding, replying and deleting are not available to automation.
The record your auditor will want.
- Every AI run recordedThe model, configuration version, cost and correlation id of every agent run, in your own database.
- Prompts and responses keptEach model call's prompt and response, hashed and linked to the run that made it, so any output can be reconstructed.
- Every human decision recordedWho accepted, overrode or rejected an AI output, and who reviewed it, with every earlier version of the record kept.
- Configuration changes reviewedA change to a prompt, model or setting reaches production only through a reviewed pipeline that records who ran it and why.
Model risk and integrations.
- Each agent pinned to a named modelEvery agent runs on a named model deployment with a versioned configuration, checked in CI. Which models run our agents is published.
- Impact assessments kept currentOur ISO/IEC 42001 AI impact assessments are generated from each agent's live configuration, and a build fails when they drift from it.
- Models reached through AzureModels are reached through our own Microsoft Azure AI Foundry resources. A build rule blocks direct calls to OpenAI's public API.
- A connection your IT team controlsThe Microsoft 365 app your IT team approves can read and file mail and write drafts. It does not ask for permission to send mail, and your team can see and remove it.
- Your credentials kept apartCredentials you give us, such as connector tokens and API keys, are held in a separate key vault from our own secrets. Our web application can write them but cannot read them back.
Audits and assurance.
- SOC 2 Type II completeThe report is available on request through our trust centre. Infrastructure evidence is collected from live Azure for each audit period.
- Encrypted at rest and in transitTransparent Data Encryption on every production database, and TLS in transit.
- ISO/IEC 42001 programme under wayOur AI management system is being built to the international standard for governing AI. Where the programme stands.
Questions reviewers ask.
Can the AI send email or act without us?
No. Lynctera never sends email for you: replies are saved to your Drafts folder and you send them from Outlook. Filing rules your team sets up can move, flag and archive mail, never forward, reply or delete it. A step that needs approval waits for the person it is assigned to.
Where is our data held?
In your firm's own Azure SQL database in the UK, separate from every other customer, and encrypted at rest and in transit.
What does the architecture review cover?
An engineer walks your team through tenant isolation, human sign-off, the audit trail and how model changes are made, in the code and configuration that enforce them, and answers your questions as you go. It takes about 45 minutes.
Will you complete our security questionnaire?
Yes. Send it with your request. Our SOC 2 Type II report and policies are available through our trust centre.
How do we report a security concern?
Email support@lynctera.com with “Security” in the subject line. We handle these ahead of the queue and will tell you what we found.
See the evidence for yourself.
Book a 45-minute architecture review with an engineer. Bring your questionnaire.