Security
Security built for regulated work.
FINTERRA agents run inside your environment, under your controls. Models never make regulated decisions, every action is logged, and your security team can inspect what they deploy.
AI controls
The model assists. It never decides.
Agents are built so that a model cannot take a regulated decision, move money or contact a customer outside the rules you set.
Deterministic decisions
Eligibility, pricing, scoring and calculations run in your rules engine. The model prepares inputs and explains outcomes.
Human approval gates
Credit, claims, consumer-contact and advice decisions route to a named person. Approval points are listed on every agent page.
Adversarial testing
Prompt injection, jailbreak attempts and fair-lending probes are part of the L4 test suite and re-run on every release.
No training on your data
Your data is never used to train models - ours or a provider’s. Agents use your data at runtime only.
Bring your own model
Run agents on the model provider your organisation has approved, with your own keys and your own data-processing terms.
Prompts under change control
Prompts and policies are versioned with the agent. Any change is re-tested against its suite before release.
Access control
Least privilege, by declaration.
Every tool an agent can call is declared up front with its access level. Nothing is reachable that is not in the specification.
Declared read / write
Each tool is marked read, write or both. Write tools are few, scoped and usually behind an approval step.
Scoped service identities
Each MCP server connects with its own service identity, limited to the endpoints and records its tools need.
Your identity provider
Staff access to agent workbenches and approval queues goes through your SSO, MFA and role model.
Your secrets manager
Credentials live in your vault and are injected at runtime. They are never stored in agent images or configuration.
Data handling
Your data stays in your environment.
Agents and MCP servers run as containers in your cloud account or data centre, so customer data stays under your existing controls.
Runs in your environment
Deploy into your VPC, private cloud or on-premises cluster. No customer data flows to us.
Encryption
TLS for every connection between agent, MCP servers and your systems. Data at rest uses your storage and your keys.
Minimisation
Agents request only the fields a step needs and hold working context for the task, not beyond it.
PII-aware logging
Sensitive fields are redacted in logs by default. What is retained, and for how long, follows your policy.
Logging & monitoring
Every step on the record.
Examiners and internal audit can reconstruct what an agent did, why, and who approved it.
Complete audit trail
Each tool call, input, rule version, model output and human decision is logged with a timestamp and case reference.
Into your SIEM
Logs and metrics export to your SIEM and observability stack in standard formats.
Drift monitoring
Accuracy and outcome metrics are tracked in production, with alerts when an agent moves outside its tested range.
Software supply chain
Know what you are running.
Every release is built, scanned and versioned so your security team can review it before it enters your environment.
Signed images
Container images are signed, so you can verify that what you deploy is what we released.
Software bill of materials
Each release ships with an SBOM listing its components and versions.
Vulnerability scanning
Images and dependencies are scanned before release, and patched releases are issued for critical findings.
Re-tested releases
No version ships without passing the agent’s full test suite, from L1 tool use to L4 compliance.
Regulatory compliance
Evidence your risk team can use.
Each agent maps the regulations it touches - ECOA, FCRA, Reg F, UDAAP, GLBA and more - to the guardrail or control that addresses them, and ships with model-risk documentation for your review.
Browse agents by regulationResponsible disclosure
Found a vulnerability?
Email [email protected] with the details and steps to reproduce. We acknowledge reports quickly, keep you updated and credit researchers who want it. Please don’t access customer data or disrupt services while testing.
Questions
What security teams ask us.
Does any of our customer data leave our environment?
No customer data comes to us. Agents and MCP servers run inside your environment. Model calls go to the provider you choose, under your agreement with them, or to a model you host yourselves.
Is our data used to train models?
No. Neither we nor your model provider (under its enterprise terms) train on data processed by your agents.
Can we review an agent before it goes live?
Yes. Every agent has a public specification - workflow, tools, guardrails, approval points and test levels - and your teams receive the full documentation pack and test suite. See how it works.
Who can see or change prompts and policies?
Prompts are not published. Inside your deployment, policies, thresholds and templates are configuration owned by your teams, under your change control, and re-validated before release.
How do you handle a vulnerability in an agent or MCP server?
We triage reported and detected issues, ship a patched, re-tested release and notify affected customers with the details your security team needs.
Can we get your security documentation?
Yes. Request it through the demo form and we will share our security pack and architecture documentation under NDA.
Bring your security team to the demo.
We’ll walk through the architecture, controls and documentation pack.