Built for the accreditation pack you actually have to pass.
Not a security checklist bolted onto a SaaS agent. Udal is engineered for regulated finance and government from the first commit — self-hosted, fail-closed, and provable end to end.
Never leaves your boundary.
Every component runs inside your environment. Frontier models are reached in-tenant. Open-weights models serve from your own GPUs. There is no vendor pipe to be breached, subpoenaed, or repriced.
Can't be talked into breaking the rules.
A central control plane screens every model call and tool invocation before it happens. Agents inherit the invoking user's permissions, not a shared service account. High-impact actions require a different human to approve.
Gets an answer, not a shrug.
Every decision, model call, tool use and approval is written to a tamper-evident audit chain and mapped to SOC 2, ISO 27001, NIST 800-53. Exported as a sealed report an assessor can trace control by control.
§ 03 — First principle
"If any control is missing or misconfigured, the platform refuses to start and names the exact gap."
One switch. Fail-closed. No silent insecure defaults in production.
Answered up front. In writing.
?>How do you know an agent didn't leak PII?
·The gateway redacts it before the model sees it. If the classifier is missing, the platform refuses to start.
?>How do you stop an agent burning through budget?
·An admin sets a per-project hard cap. It overrides the agent's own manifest. The author cannot loosen it.
?>How do you prove the audit log wasn't edited?
·It's HMAC hash-chained and append-only. Any gap is mathematically detectable, and must be re-verified after any restore.
?>How do you stop someone deploying an unsigned image?
·A Kyverno admission policy in your cluster rejects any Udal image not signed by the release key. Enforced at run, not audited after.
?>How do you run this in an air-gapped network?
·One signed bundle. Offline license verification. Pre-staged model weights. No outbound telemetry. No exceptions.
Evidence,
not marketing badges.
Udal generates a sealed, control-mapped evidence report from its own audit chain, RBAC, policy and change records — the artefact your assessor actually wants. No hand-assembled spreadsheet.
The full pack — threat model, control mappings, pen-test summary — is shared under NDA on request.
See how it's deployed on your ground.
Run governed agents on ground you own.
Technical briefing, not a sales deck. A human replies within one working day.
Request access