Key facts
| Deployment models | Plugsky cloud, private endpoint in your VPC, on-prem, air-gapped |
| Audit events | Authentication, key lifecycle, admin and policy changes, usage metadata |
| Audit export | Stream or export to SIEM (Splunk, Sentinel, QRadar, Datadog) |
| Prompt logging | Configurable; disable or minimize prompt retention for sensitive workloads |
| Retention | Set per deployment; contractual details in the DPA and SLA |
| Encryption | AES-256 at rest, TLS 1.3 in transit; BYOK via KMS or HSM |
| Compliance posture | SOC 2 Type II and ISO 27001 readiness in progress (not yet certified); GDPR and PDPL alignment |
| Regions | EU (Frankfurt), GCC (UAE), APAC (Singapore), US (Virginia) region-locked planes |
TL;DR
- Log identity events, key lifecycle, model invocations, and every administrative or policy change.
- Metadata is usually enough: actor, action, resource, model, token counts, timestamp, source IP.
- Minimize or disable prompt retention for sensitive workloads, and document that decision.
- Make logs tamper-evident and export them to your SIEM — internal dashboards are not evidence.
- Map each log category to the control it supports so audits become a lookup, not an investigation.
How it works, step by step
- List the frameworks you must satisfy (for example GDPR, PDPL, SOC 2, internal policy) and the evidence each requires.
- Define the event schema: actor, action, resource, model, tokens, outcome, timestamp, source.
- Decide prompt and completion retention per workload class; disable it where it is not needed.
- Configure streaming or scheduled export to your SIEM and test schema stability.
- Set retention windows and access controls for the log store itself.
- Run a tabletop exercise: reconstruct one user session from logs alone.
- Review coverage quarterly as new models, tools and regions are added.
Try it yourself
Open the sovereign AI readiness score →
Events worth logging
- Authentication and session — sign-in, SSO assertion, failed attempts, session termination.
- Key lifecycle — creation, scope changes, rotation, revocation, last-used metadata.
- Inference — model ID, endpoint, token counts, status, latency, workspace and key used.
- Administration — role changes, policy edits, webhook and integration changes.
- Data actions — uploads, vector-store writes, exports, deletions.
Log metadata rather than raw content wherever possible. Content logging multiplies privacy and residency obligations without improving incident response. If an event cannot be tied to a control you operate, stop logging it — volume without purpose increases cost and exposure.
Retention, immutability and residency
Retention should follow the shortest period that satisfies your policy and local law. Store logs in the same residency plane as the workload, encrypt them, and restrict access to a small audit team. Tamper-evidence matters: a log an administrator can silently edit is weak evidence. Plugsky supports audit export to standard SIEM tooling and region-locked data planes so logs do not silently leave the jurisdiction you selected.
Common pitfalls
- Logging prompts by default, then discovering the data includes regulated personal data.
- Retaining everything forever with no deletion path.
- Exporting to a SIEM but never validating that fields arrive intact.
- No logs for administrative changes, which are the first thing an auditor samples.
- Assuming a vendor dashboard export satisfies legal hold requirements.
How to use this in procurement
Ask vendors for a sample event, the export format, retention options, and where logs are stored. If the answer is a dashboard screenshot rather than a schema and an endpoint, treat audit capability as unproven. Plugsky documents SIEM export and region-locked storage; validate the specifics in a trial against your own schema before signing.
Honest comparison
| Capability | Plugsky | Hyperscaler AI platform | Building in-house |
|---|---|---|---|
| Audit scope | Identity, keys, admin changes, inference metadata | Broad cloud-native audit coverage | Depends entirely on your code |
| SIEM export | Splunk, Sentinel, QRadar, Datadog | Native cloud logging and connectors | You build and maintain pipelines |
| Prompt retention control | Configurable and minimizable per workload | Varies by service | You define and enforce it |
| Residency of logs | Region-locked planes (EU, GCC, APAC, US) | Region selection in most services | Wherever you host |
| Tamper evidence | Managed platform controls | Mature cloud controls | Your responsibility |
Frequently asked questions
Do we need to log prompts and completions?
Usually no. Metadata answers most incident and audit questions with far less privacy exposure. Log content only where a specific obligation requires it, and document the decision.
How long should AI audit logs be retained?
Match your sector and framework, then keep the shortest defensible window. Common baselines are 90 days to 12 months; regulated sectors may require longer. Confirm against local law.
Can Plugsky send logs to our SIEM?
Yes — audit log export to Splunk, Sentinel, QRadar and Datadog is documented. Test the event schema during a trial.
What should an auditor expect to see?
Evidence that access is provisioned and revoked deliberately, that model usage is attributable to a key and workspace, and that administrative changes are recorded.
Do audit logs leave our region?
Logs stay in the residency plane you select for the workspace. Confirm the exact region mapping and any subprocessor involvement in writing.
How does this relate to data residency?
Residency governs where data is stored and processed; audit logs record access to it. Both must point at the same jurisdiction for a clean compliance story.
Is logging required for the free plan?
The free plan is for evaluation. Enterprise agreements define retention, export and residency requirements — discuss them when you move to production.