A patched stack, from firmware to model runtime.
A private AI environment has more layers to patch than a typical application: firmware, OS, GPU drivers, container runtime, serving software, databases, and Control itself. We track vulnerabilities across all of them and apply updates on a schedule you approve.
Everything that needs patching.
Compatibility between layers matters: a driver update can break a serving runtime. We test combinations before applying them.
| Layer | Examples | Notes |
|---|---|---|
| Firmware | BMC, BIOS, GPU firmware | Vendor tooling; often needs a reboot window |
| Operating system | Kernel, packages | Coordinated with driver versions |
| GPU stack | Drivers, CUDA libraries | Must match serving runtime support |
| Containers | Runtime, base images | Rebuilt and rescanned |
| Serving | vLLM, Ollama | Re-validated against registered models |
| Data stores | PostgreSQL, pgvector | Minor versions routinely; majors planned |
| Data.co Control | Signed releases | Applied from your mirror or offline media |
How a patch reaches production.
- 1
Assess
Rate each advisory for relevance and severity in your environment.
- 2
Test
Apply in a staging environment and run validation, including model serving checks.
- 3
Schedule
Apply in your approved change window, or an emergency window for critical issues.
- 4
Verify
Confirm versions, run health checks, and record the change.
Hardening and review.
Configuration baselines
Hosts and services checked against an agreed hardening baseline.
Access reviews
Regular review of administrative accounts, including Data.co's own access.
Secret rotation
Credentials and certificates rotated on schedule and before expiry.
Exception tracking
Anything that cannot be patched is documented with rationale and compensating controls.
A rhythm, not a scramble.
Routine updates are batched into a recurring change window that your team approves once and then sees on a calendar. Each window has a written plan: which layers move, which versions, the staging results, and the rollback for each step. Firmware and kernel changes that need reboots are grouped so nodes cycle once rather than several times a month.
Out-of-band changes follow a separate, pre-agreed path. When an advisory is rated critical for your environment, we notify your contact with the assessment, the proposed fix, and any interim mitigation, such as disabling an exposed interface, and use the emergency window defined in the contract. Severity ratings use the vendor's score as an input, but the decision weighs exposure in your network: a flaw in a service that is not reachable is handled differently from one on the gateway.
Data.co Control releases arrive as signed images, and policy changes as signed bundles. The runtime verifies signatures and runs policy tests before activating anything; a failing test keeps the previous policy set in force. That gives patching a safe failure mode at the governance layer, which matters when an update window is short.
Security duties under a managed contract.
| Duty | Data.co | Your team |
|---|---|---|
| Vulnerability tracking across the stack | Monitors advisories and scans | Receives assessments |
| Patch testing and deployment | Stages, applies, verifies | Approves windows and emergencies |
| Hardening baseline | Proposes and checks against it | Approves the baseline |
| Identity provider and MFA | Integrates Control with it | Operates and secures it |
| Policy signing keys | Never holds them by default | Keeps them offline and signs bundles |
| Incident response | Contains within agreed authority, preserves evidence | Leads the response and notifications |
Records your auditors can use.
Patch levels, change records, exceptions, and access reviews are reported regularly, designed to support your internal audit and compliance processes.
Questions buyers ask
How fast are critical vulnerabilities patched?
Timelines by severity are defined in your contract. Critical issues can use emergency change windows agreed in advance.
Will patching cause downtime?
Some updates, such as firmware and kernel patches, need reboots. With redundant nodes, these can be rolled through without full outages; single-node deployments need a maintenance window.
How do you patch air-gapped environments?
Through signed offline media under your transfer procedure, with the same testing and records, coordinated with on-site staff.
Does this satisfy a specific security framework on its own?
No service alone does that. This work is designed to support control requirements common to security frameworks, and the records help your compliance team show it.
Who reviews Data.co's own access?
Your team, with our help. Access reviews list every account our engineers use, what it can reach, and when it was last used, so you can remove anything that is no longer needed.
What happens if a patch breaks model serving?
Each change has a rollback step written before the window opens. Serving checks run immediately after the change, and if they fail we roll back and record the result before trying again.
See it against your own use case.
A 30-minute technical walkthrough, policy simulation on your scenarios, and a deployment recommendation for on-prem, private cloud, or hybrid.