Security and evaluation
Security and testing are part of how we build, from the first release. These are the practices we follow on every project, and you can hold us to them.
Led by an Army veteran. Built by engineers who have shipped production systems for NASA and Cigna, and at Amazon Web Services (AWS), Bloomberg, and Capital One.
AI systems
Most AI projects fail on cost, risk, or unclear value rather than on the model itself. We spend our effort on the controls around the model.
We build a test set from your real cases, agree on the accuracy bar before results are in, and keep measuring after launch. A made-up source counts as a failure.
Each agent runs in an isolated environment with only the tools, data, and permissions its task needs. It can't reach anything else.
Low-risk steps run on their own. Anything consequential, like sending, paying, or changing a record, waits for a person. You decide where that line sits.
Every question, source, draft, approval, and action is logged. Unusual behavior is flagged, and any agent can be stopped immediately.
Spending caps and cost tracking for each workflow, so running costs are known before you scale.
The full approach is on the AI page: build, evaluate, and govern
All software
No change reaches production without a second set of eyes, whether a person or an AI wrote the first draft.
Tests run on every change, so a fix in one place doesn't quietly break another.
Every build is checked for known vulnerabilities, and each release ships with a software bill of materials (SBOM).
Source code, infrastructure, documentation, and data belong to you from day one.
Security and data
We design to the NIST AI Risk Management Framework and to the security controls your authorization (ATO) review will ask for.
Systems run in your cloud account, GovCloud, or on your own servers.
People and agents get the minimum access the job requires.
Data is encrypted in transit and at rest, with keys you control.
Your data is never used to train AI models.
Our commitments
These are written into how we work, so you can check them on any project.
Ship an AI system without a test set and an agreed accuracy bar.
Let an agent take a consequential action without a way for a person to approve or stop it.
Use your data to train AI models, or move it outside your environment without your approval.
Recommend AI when plain software would do the job better.
Contact
Send a short note: what the system does today, what isn't working, and any deadlines. An engineer will reply within two business days.
contact@autonar.ai