In-VPC deployment
Run the whole Sentisec platform inside your own network boundary — what sits where, what egress it needs, and how to provision it.
In-VPC is the deployment shape for regulated workloads. The entire platform runs inside your own network: your agents' prompts, the verdicts, and the evidence records never leave it, and no Sentisec-operated system can see any of them.
This page is what to hand your platform team. The deployment package itself — images, configuration reference, and the runbooks — comes with the engagement.
What runs inside your boundary
| Component | Inside your VPC |
|---|---|
| The API (request path + processing) | yes |
| Operator dashboard | yes |
| Documentation | yes |
| Database | yes — managed Postgres is fine |
| Cache / event bus | yes — managed Redis is fine |
| Object store for evidence records | yes — S3, GCS, or Blob |
| The key that encrypts workspace credentials | yes — your own KMS or secrets manager |
| Evidence records | yes — your bucket, your retention |
Nothing in that list phones home. There is no Sentisec-side telemetry channel, no shared database, and no path by which we could read a session even if asked to.
What has to be reachable from it
Only the model providers you actually connect. Sentisec's outbound requirement is exactly the set of upstream endpoints your workspaces use, and nothing else:
| If you connect | Reachable host |
|---|---|
| Anthropic | api.anthropic.com |
| OpenAI | api.openai.com |
| Kimi (Moonshot) | api.moonshot.ai |
| Gemini | generativelanguage.googleapis.com |
| Vertex AI | <region>-aiplatform.googleapis.com, in your own project |
Vertex is the interesting case for a regulated deployment: the endpoint lives in your Google Cloud project and region, so the traffic can stay inside your cloud's own network via Private Service Connect rather than traversing the public internet at all.
Identity. By default the dashboard authenticates through Sentisec's managed identity layer, which needs outbound HTTPS. If that is unacceptable — sovereignty requirements, or a genuinely disconnected environment — a self-managed install can instead authenticate against your own OIDC provider directly, with no Sentisec-operated service anywhere in the sign-in path. See the self-managed OIDC guide.
Provisioning checklist
Network. One private subnet per availability zone. Public subnets are not required — TLS terminates at an internal load balancer.
Encryption key. Create a customer-managed key for the workspace credential encryption root. The Sentisec workload needs decrypt permission only; it never needs to encrypt once the key is bootstrapped.
Database. Managed Postgres 16 or later. TLS-only connections, encrypted at rest with your key, automated backups, point-in-time recovery.
Cache. Managed Redis 7 or later. TLS-only, encrypted at rest, persistence enabled, password from your secrets store.
Object store. A bucket for evidence records, encrypted with your key. Scope the bucket policy so only the Sentisec workload identity can read or write objects.
Secrets. Provision in your secrets manager: the workspace encryption root key, the database password, the cache password, and your identity-provider configuration.
TLS. Internal load balancer with a corporate-PKI or cloud-managed certificate. Give the API, the dashboard, and the docs their own internal hostnames.
Egress. Restrict outbound traffic at the network layer to the provider hosts above. Sentisec's own egress allowlist is a second, per-workspace layer inside that boundary — set both.
Deploy, using the package and the configuration values from your engagement.
Smoke test. Connect one provider credential and use the dashboard's connection test. Sign in through your identity provider. Run one agent session end to end and confirm it appears in the session list with a verdict.
Compliance posture
- SOC 2 CC6 — your own key management, encrypted at rest and in transit, an append-only audit log, signed evidence records, and role-based access.
- NIST AI RMF — Govern: what each workspace enforces, plus roles. Map: the threat surface, visible per session. Measure: verdict and latency metrics per workspace. Manage: evidence records, alert rules, and SIEM forwarding.
- EU AI Act Article 12 — content-hashed, signed evidence records with a retention window you set, tagged to the control they evidence.
Related
- Deployment options — how in-VPC compares
- Self-managed OIDC — sign-in with no external dependency
- Log exports — getting findings out
- Request the deployment package