Security & trust
Security from the platform up.
Built into the platform, not bolted on after. The controls below run in production for every customer, and we maintain them so your team does not have to.
Posture
Where we stand, stated plainly.
NIST SP 800-171 · CMMC 2.0
Our controls are mapped to NIST SP 800-171 Rev. 2 and the CMMC 2.0 Level 2 practice set, and we maintain the system security plan and plan of action that assessors expect. We say "aligned to" until a third party says otherwise.
Deployed inside your boundary
The runtime lives in your accredited environment: AWS GovCloud or Azure Government at IL4/IL5, your on-prem enclave, or a disconnected network. Your authorizing official stays the authorizing official.
Secure by design
Safeguards are part of the architecture, not configured per project or sold as add-ons. Every customer gets them from the first day, and nothing depends on someone remembering.
Architecture
Everything runs inside your boundary.
Data stays where it lives. Models come to it. The only things that cross the line are a VPN session and a signed release.
Scroll sideways to see the full diagram
Platform safeguards
Six controls that are part of the architecture.
Government cloud or your metal
FedRAMP High regions in AWS GovCloud and Azure Government, with private endpoints, a web application firewall and no public ingress. On-prem and air-gapped deployments run open-weight models on your GPUs.
Tenant isolation
Each customer gets its own data stores, identities, access controls and message queues. Production is separated from development, test and staging with no shared credentials.
Access control
Your identity provider, including CAC/PIV through SAML or OIDC. Role-based access at component and record level, least privilege, MFA for everyone, session timeouts and quarterly access reviews.
Change governance
Automated tests and mandatory peer review before anything reaches production. Engineers cannot deploy their own changes. Builds are signed and reproducible.
Full audit trails
Immutable logs of data operations, access changes, model calls and releases, exportable to your SIEM. Records are versioned with rollback.
Model governance
Models selected by data sensitivity. Low-confidence outputs go to a person. Reasoning is logged, drift triggers review, and autonomy is set per use case according to risk.
People & process
The people who touch your systems.
- U.S. persons and clearances
- Engineering staff on defense work are U.S. persons. Cleared personnel are assigned where the program requires it, and we will discuss facility and personnel arrangements for classified work case by case.
- Background checks and offboarding
- Everyone with access to customer systems is background-checked first, including contractors. Access is revoked within one business day of departure or role change.
- Training and rules for AI tools
- Security awareness training before any customer data access, and an acceptable use policy that covers systems, customer data and the AI tools our own staff use.
For your security review
The questions diligence asks.
- Data residency
- Data stays in the region and environment you choose at the start. In disconnected deployments it never leaves the room.
- Export control
- We design for ITAR-controlled data: U.S. persons only, U.S. regions only, and no commercial model endpoint in the path unless you approve it in writing.
- Model providers
- Open-weight models on your infrastructure; hosted models in GovCloud regions; commercial APIs only by explicit opt-in for unclassified data. All arrangements exclude your data from model training.
- Encryption
- AES-256 at rest with keys you can hold, TLS 1.2 or higher in transit, FIPS 140-validated modules where the environment supports them, secrets in a per-tenant vault with rotation.
- Retention and deletion
- Retention set per record class to match your schedule. Deletion is permanent. On exit, data and configuration are returned and the environment is scrubbed and attested.
- Vulnerability management
- Continuous posture assessment, dependency and container scanning before every build, and remediation timelines by severity written into the agreement.
- Incident response
- Continuous monitoring with incidents graded Low to Critical. High and Critical go straight to the information security officer. Notification windows are set in the contract, and DFARS 252.204-7012 reporting is supported.
Common concerns
Addressed.
- Do you train models on our data?
- No. Provider agreements exclude customer data from training, and our own models are tuned only on your data for your solution, which you own.
- Can we choose where models run?
- Yes. On your hardware, in GovCloud, or through a commercial endpoint for unclassified work, set per workflow.
- Who can access our environment?
- Only named, background-checked U.S. persons, through your identity provider, with MFA, under role-based access, with every action logged.
- What happens to our data if we stop?
- All data, schemas, workflows and configuration are returned in open formats, then the environment is decommissioned and the deletion attested.
- Can we get security documentation?
- Yes. The policy framework, the NIST SP 800-171 control mapping, the system security plan and architecture diagrams are available under NDA on request.
Ask us the hard questions
Bring your assessor. Bring your ISSO.
Any topic on this page, in as much depth as your review needs, with the engineers who built it in the room.