Security
Security, layer by layer.
How we build AI systems that can hold classified, protected and personal data — from the network up to the model itself.
Architecture
Contained by design. Every client environment is separate, encrypted with keys the client holds, and reachable only over private networks.
Zero-trust network
No implicit trust between services. Every call is authenticated with short-lived credentials and authorised against policy.
Isolation per client
Each client runs in its own account or enclave — separate compute, storage, keys and logs. Nothing is multi-tenant by default.
Client-held keys
Data encrypted at rest with keys in the client’s custody. Revoking a key makes the data unreadable, including to us.
Private connectivity
Model endpoints are reachable only over private networking or VPN. There is no public inference URL to attack.
AI-specific threats
The attacks a firewall does not see. Language models add new ways in. We design against them explicitly rather than trusting the model to behave.
Prompt injection
Retrieved documents are treated as data, never instructions. Tool calls are allow-listed and every action a model can take is scoped to the user.
Data exfiltration
Outputs are filtered for sensitive patterns, and retrieval is permission-checked before the model sees a passage — not after.
Training-data poisoning
Training sets are versioned, sourced and reviewed. A model can be traced back to the exact data snapshot that produced it.
Model supply chain
Base models and libraries come from verified sources, are scanned on intake and pinned by hash before deployment.
Engineering
Security in the pipeline. Controls run automatically on every change, so they hold on the hundredth release as well as the first.
Secure development
Threat modelling at design, mandatory peer review, and security requirements written into every work item.
Automated scanning
Static analysis, dependency and container scanning, and secrets detection on every commit, blocking on high severity.
Independent testing
Third-party penetration testing at least annually and before major releases, including AI-specific test cases.
Reproducible releases
Infrastructure as code and signed build artefacts, so what was reviewed is exactly what is running.
People & operations
Who can touch it, and when. Most breaches are access, not exploits. Access here is narrow, temporary and recorded.
Personnel security
Background checks for all staff, and security clearances at the level each client environment requires.
Least privilege
Production access is just-in-time, approved per request, time-boxed and logged.
Incident response
A tested plan with named roles. Clients are notified of any incident affecting their data within 24 hours of confirmation.
Vulnerability disclosure
Researchers can report issues to security@ainvento.com. We acknowledge within two business days.
Need it in writing?
The security brief covers all of this in the form an evaluation team files. The full package, including audit reports, is available under NDA.