Skip to content
Vraelis
DocsPricing
Talk AI security
Network servers and cabling in a data center.

Model integrity

Know which model is running.

Model files, configurations and dependencies can change between development and deployment. We are investigating how to establish provenance, verify approved artifacts and make changes reviewable.

Talk AI security→→Development status

The work

Integrity begins before deployment.

An authentic artifact can still behave unsafely. Provenance, tamper checks and behavioral evaluation answer different questions.

Provenance

Establish the source

Record where an artifact came from, how it was produced and which dependencies it includes. A version label supplied by the workload is insufficient evidence.

Verification

Compare the approved artifact

Bind deployment to a reviewed model and configuration. Signatures and hashes can establish authenticity and change; they cannot establish freedom from backdoors or adversarial vulnerabilities.

Change

Review what changed

Preserve the relationship between an update, its authorization and the evidence used to evaluate it. Model weights, preprocessing and runtime configuration all affect behavior.

A microcontroller and tools on a work surface.
Illustrative photography · Sources

Deployment context

The artifact is part of a running system.

A verified model file does not establish which workload has loaded it or whether the surrounding software is intact. Runtime identity and trustworthy measurements need a defined source and integration.

Model integrity is a development priority. We do not offer a deployed attestation service, universal model scanner or verified support for specific edge hardware.

What comes next

Test the integrity boundary.

Define the model format, runtime and threat before choosing the verification mechanism.

Baseline

Identify the reviewed configuration

Name the artifact, preprocessing, dependencies and intended environment. Document what the available evidence can authenticate.

Evaluation

Exercise unauthorized changes

Use controlled tests to check altered artifacts, mismatched configurations and invalid provenance. Record which changes remain outside coverage.

Operations

Preserve independent recovery

Integrate update and recovery procedures with the system owner. Verification failures need a system-specific response that respects existing operating controls.

Talk AI securitySecurity problemsZero trust
Vraelis

AI security. In development.

Defense. Infrastructure. Robotics.

Product

PlatformZero trustRecorded evidenceDevelopment statusIntegrationsAI workloadsPricing

Solutions

DefenseInfrastructureRoboticsGovernment & institutionsSystem integratorsEnterpriseExplore solutions

Resources

DocumentationDeveloper toolsThe problemsResearchChangelog

Company

Our goalsAboutContact

Trust

SecurityPrivacyCookiesTermsAcceptable use
More policies
LimitationsRefundsData rightsSubprocessorsTrademark
Security for AI in the physical world.

Developing cybersecurity for AI-enabled defense, infrastructure and physical systems.

© 2026 Vraelis
Image sourcesLinkedIn