Short answer. A SOC 2 logo is not evidence. Ask for the report, its period, its trust services criteria and the auditor, then test the answers against your own risk model. For AI, add four questions the standard questionnaire misses: which subprocessors see prompts, what is logged, who holds keys, and how fast the vendor notifies you. Contract terms, not marketing pages, decide the rest.
Quick answers
The table below is a starting point for a first review. Each answer points to the document or clause that should settle the question.
| Buyer question | One-sentence answer |
|---|---|
| Is Plugsky SOC 2 certified? | No public certification claim is made; SOC 2 is described as readiness or in progress, so request the current report under NDA and verify its period. |
| Does Plugsky sign a DPA? | Yes — GDPR Article 28 terms and EU SCCs are available on paid plans, with bespoke addenda on Enterprise, subject to review of the actual document. |
| Can we get a subprocessor list? | Ask for the current list, locations, change notice and objection rights; do not infer it from the model catalogue. |
| Is HIPAA covered? | A BAA is available by contract, subject to review, and there is no HIPAA certification. |
| What uptime is promised? | A 99.9% headline target, with plan-level terms and exclusions in the contract; Hobby is not in the SLA table. |
| Can we bring our own keys? | Key custody is contract-led; read the BYOK page and confirm the key hierarchy in writing. |
| How fast is breach notification? | No public window is specified, so negotiate the clock, severity levels and escalation path in the DPA. |
| What should we test first? | Test SSO enforcement, log export, deletion of a test dataset and one incident tabletop before you scale. |
SOC 2 readiness is not a SOC 2 report
“SOC 2 ready” and “SOC 2 examined” sound similar in a sales deck and mean very different things in a risk register. The first is a project; the second is evidence from an independent CPA firm.
Is Plugsky SOC 2 certified?
Plugsky does not claim SOC 2 certification, and buyers should not accept a logo as proof. Public material describes SOC 2 as readiness or in progress. That means the control environment may be designed and evidenced internally, but it is not the same as an independent Type II opinion covering an observation period. Ask for the current report or attestation under NDA, confirm the auditor, the trust services criteria, the period covered and whether any exceptions were noted. If no report exists yet, ask for the readiness assessment and the remediation dates. The security review page is the starting point, not the evidence.
What is the difference between SOC 2 readiness and certification?
Readiness is an internal or consultant-led gap assessment against the AICPA Trust Services Criteria. A SOC 2 examination is performed by an independent CPA firm and produces a report: Type I covers design at a point in time; Type II covers design and operating effectiveness over a period, typically three to twelve months. There is no single “SOC 2 certificate” issued by the AICPA. A vendor can be ready, in progress or examined, and those are different risk positions. Ask which one you are looking at, and read the scope and exceptions, not just the cover page.
Which trust services criteria should we ask for?
Security is the common criterion; availability, confidentiality, processing integrity and privacy are optional. For an AI API, security and availability are the baseline. Confidentiality matters if prompts contain commercial secrets or personal data. Privacy is relevant only if the vendor processes personal data as a controller, which is usually not the case for an API processor. Ask for the criteria in scope, the systems and services covered, and whether the report includes the model-serving environment or only corporate IT. A report that excludes the production inference stack answers the wrong question.
Can we see the report before signing an NDA?
Most vendors will not release a SOC 2 report without an NDA, and that is reasonable. What should not wait for signature is the report metadata: report type, period, auditor, criteria, and whether it is a readiness assessment or an examination. Ask for a bridge letter if the period ended more than a few months ago, and ask what changed since the period closed. If the vendor cannot state the period or the auditor, treat the SOC 2 claim as marketing until proven otherwise. Put the report request in the DPA review checklist so it is tracked.
DPA, transfers and subprocessors
The data-processing agreement is where the commercial promise becomes an enforceable obligation. Read it with the data-flow diagram open.
Does Plugsky offer a DPA and EU SCCs?
Yes. A data processing agreement (DPA) with GDPR Article 28 terms and the European Commission's Standard Contractual Clauses is available on paid plans, with bespoke addenda on Enterprise. That is the starting point, not the end of review. Check the controller and processor roles, the instructions for processing, deletion and return terms, audit rights, liability caps, subprocessor notice and the transfer mechanism. If your data is Saudi, Emirati or DIFC/ADGM-regulated, confirm that the addendum you sign matches the regime you are actually relying on. Counsel should review the executed version, not a sample.
How do we review subprocessors, including the model provider?
Ask for the current subprocessor list with locations and roles, the notice period for changes, and the objection right. For an AI API, the model provider is the obvious dependency, but hosting, observability, support and payment providers may also touch data. Do not infer the contractual list from a model catalogue: a company can be an upstream model licensor without being a subprocessor for your prompts, or the reverse. Plugsky's model ladder lists NVIDIA as the upstream provider for most public model IDs; whether that makes NVIDIA a subprocessor for your data is a contract question, so ask Plugsky to state it in writing.
What happens to data at contract termination?
The DPA should state the deletion window, the backup exception, the format for return, and the certification you receive at the end. “Deleted on request” is weaker than “deleted within 30 days and backups purged within 90”. For AI workloads, ask separately about prompts, completions, logs, embeddings and fine-tuned artifacts; these are often stored in different systems with different retention. Also ask whether derived data, such as evaluation sets or abuse-monitoring samples, survives termination. Get the answer in the schedule, not in a sales call.
How do we handle a transfer impact assessment?
Map the data flow first: where the prompt is sent, where it is processed, where logs and backups live, and which support staff can access them. Then identify the transfer mechanism — adequacy, SCCs or another Article 45/46 route — and assess the destination's legal environment and supplementary measures. For AI, the model provider's location and any remote support access belong in the map. The European Commission's SCCs do not replace the assessment; they are one input. Your counsel and DPO should sign off the conclusion, especially for government, health or financial data.
Incident response, audit logs and keys
These are the controls that decide how quickly you can answer a regulator, and how much you can reconstruct after an incident.
What should an AI vendor's incident response clause contain?
An AI vendor's incident response clause should include, at minimum, a severity taxonomy, a notification window that starts when the vendor becomes aware, a named contact and escalation path, regular updates, a root-cause report, and cooperation with your regulators and customers. For AI, add prompt-injection or model-abuse incidents, training-data leakage, and unauthorized model or endpoint changes. Ask whether the clock runs from detection or confirmation, and whether the vendor can notify you before its own public disclosure. No specific window is published for Plugsky, so negotiate it rather than assuming 24 or 72 hours. Read the SLA page alongside the incident schedule.
Which audit logs should we require?
Require authentication events, key creation and revocation, model and endpoint changes, administrative role changes, data export and deletion, and per-request metadata with timestamp, model ID, token counts and status. For pooled models, response metadata can identify the serving engine so you can reconstruct which engine handled a request. Ask for retention length, export format, time synchronization, tamper evidence and whether logs are available through an API or only a dashboard. Confirm that logs do not themselves become a shadow copy of sensitive prompts.
How should SSO and admin access be scoped?
Require SAML or OIDC single sign-on, enforced for every member, with SCIM provisioning and de-provisioning if you have a directory. Separate billing, key management and production access; do not let every developer mint unrestricted keys. Require named accounts, MFA at the identity provider, short-lived sessions for admin roles, and an approval trail for production key creation. Ask whether the vendor supports multiple organizations or workspaces with isolated keys and logs. Then test one de-provisioning event end to end before you rely on the control. See SSO and SAML for the feature scope.
Who holds encryption keys, and can we bring our own?
Ask for the key hierarchy: who generates the root key, who can decrypt, where keys are stored, how rotation works, and what happens on offboarding. “Encrypted at rest” says nothing about who can read the data. If you need customer-held keys, read Plugsky's BYOK page and then confirm in writing whether BYOK applies to your deployment, whether it covers logs and backups, and whether the vendor can still support the service without key access. For high-assurance workloads, a private VPC or on-premises deployment may fit better than a shared cloud tenant.
Running the RFP without rewarding theater
A shorter questionnaire with verifiable answers beats a long one with polished prose. Score evidence, not adjectives.
What belongs in an AI security RFP?
Ask for a short, evidence-based questionnaire: data map, subprocessors, residency options, encryption and key custody, access controls, logging, vulnerability management, incident history, business continuity, deletion, model change control, evaluation data, and contract terms. Ask for artifacts, not adjectives: the report, the pen-test summary, the subprocessor list, the SLA schedule, the DPA, and a sample incident report. Keep the questionnaire to the risks that change your decision. A 300-question RFP usually produces 300 polished answers and very little evidence.
How do we score answers without rewarding marketing?
Score evidence quality separately from capability. For each control, define what “met” looks like: a dated report, a tested procedure, a named owner, or a contractual term. Give credit for a clear “not yet” with a date; penalize vague assurances. Weight the risks that matter to your regulator and your data, not the longest feature list. Keep a small set of knockout criteria, such as unresolved residency, no DPA, or no deletion commitment. Then verify two or three answers in a live session with the engineers who operate the service.
What should legal review before a pilot starts?
Legal should review, at minimum, the order form, DPA, SCCs or transfer addendum, SLA and support terms, acceptable-use policy, security addendum, subprocessor list, and the pilot's data-handling schedule. Check liability, indemnity, audit rights, suspension rights, price changes and termination assistance. For regulated data, confirm the BAA or sector addendum before the first prompt, not after. Mark every open term that depends on your regulator's approval. A pilot that processes real data without these documents creates the exact evidence problem you are trying to avoid.
What can we test in a proof of concept?
Test the controls you can observe: SSO enforcement, role separation, key rotation, log export, deletion of a test dataset, rate-limit behavior, failover, and the accuracy of the model on your own evaluation set. Ask the vendor to walk through an incident notification tabletop. Do not expect a pen test of the vendor's infrastructure; that is what the report and bridge letter are for. Document what you tested, the date, the environment and the result. That record is more useful to your auditor than another security questionnaire.
What is a reasonable exit plan?
Assume you will leave at least once. Require data export in an open format, a defined transition period, deletion certification, continued access to logs during the wind-down, and clear notice before a model or endpoint is retired. Check whether pricing or terms can change at renewal and whether you have price protection. For AI, model deprecation is a real exit event even if you stay with the vendor: an endpoint can change engines under the same public ID. Ask for the 2027 buying checklist and the deprecation policy in writing.
What to verify in a contract
These clauses decide whether the security answers survive an incident, an audit or a change of model.
- Roles and purpose. Controller, processor and subprocessor roles, permitted purposes, and instructions for processing.
- Data path. Regions for prompts, logs, backups and support access, including any remote access from outside your region.
- Subprocessors. Current list, notice period, objection right, flow-down terms and replacement rights.
- Transfers. The transfer mechanism, the assessment owner, and the governing law for each entity you contract with.
- Deletion. Separate windows for prompts, completions, logs, embeddings and backups, plus certification at exit.
- Incident response. Notification clock, severity definitions, updates, root-cause report and regulator cooperation.
- Audit rights. Report review, bridge letter, written questionnaire and an agreed path to a remote or on-site assessment.
- Keys. Key hierarchy, rotation, BYOK scope and the vendor's ability to access data without your key.
- Change control. Model, engine and endpoint change notice, evaluation rights and deprecation windows.
- Remedies. SLA credits, exclusions, termination assistance, liability and insurance.
What we did not verify
Plugsky did not provide a SOC 2 report, a pen-test summary, a subprocessor list or a signed DPA for this page. We did not test the production environment, review the incident-response runbook or confirm any certification. Plugsky describes SOC 2 as readiness or in progress and the DPA as documented; treat both as claims to verify, not facts. We also did not verify ISO 27001, FedRAMP or government accreditation, and this page makes no such claim. Where a term is described as available, the signed contract governs.
Frequently asked questions
Does Plugsky have a SOC 2 report?
Public material describes SOC 2 as readiness or in progress, not certified. Request the current report or readiness assessment under NDA and verify the period, auditor and scope.
Is a SOC 2 Type I enough?
Type I shows design at a point in time; Type II shows operating effectiveness over a period. For production, ask for Type II and a bridge letter if the period is stale.
Does Plugsky sign a BAA?
A BAA is available by contract, subject to review. It is not a certification, and the signed document defines the permitted scope.
Where does Plugsky process data?
Live regions are Frankfurt, UAE, Singapore and Virginia, with Riyadh on Enterprise. Private VPC, on-premises and air-gapped options are contract-led.
Can we audit Plugsky?
Ask for audit rights in the DPA: report review, bridge letter, written questionnaire and, for high-risk work, an agreed on-site or remote assessment.
How long are logs retained?
Plugsky does not publish a retention period. Ask for retention by log type, export format and whether prompts are stored at all.
Does Plugsky train on our data?
Do not assume either way. Ask for a contractual commitment that covers prompts, completions and embeddings, and confirm it in the DPA or addendum.
Is NVIDIA a subprocessor?
The model ladder lists NVIDIA as the upstream provider for most IDs, but the contractual role is not published. Ask Plugsky to state it in writing.
Do you support customer-managed keys?
Key custody is deployment-dependent. Read the BYOK page, then confirm the key hierarchy and whether logs and backups are covered.
What is the SLA?
A 99.9% headline uptime target. The contract schedule governs, exclusions apply, and Hobby is not in the SLA table.
Can we start with a pilot?
Yes, but define the data, the evaluation set, the exit criteria and the deletion date before the first prompt. Do not use production data without the DPA.
What is the biggest procurement mistake?
Accepting a certification logo instead of reading the report, the scope and the exceptions. The second is leaving incident notification and deletion to a side letter.
Bring the security pack to your review
Start with the documents your team can actually test, then compare deployment and pricing options.
Request security pack → See pricingSources
Primary documents we checked. Vendor claims are treated as claims; contract terms govern.
- AICPA & CIMA — SOC 2 examination overview — checked 20 November 2026.
- European Commission — Standard Contractual Clauses — checked 20 November 2026.
- EUR-Lex — Regulation (EU) 2016/679 (GDPR), Article 28 — checked 20 November 2026.
- NIST SP 800-61 Rev. 2 — incident handling guide — checked 20 November 2026.
- NIST Cybersecurity Framework 2.0 — checked 20 November 2026.
Last checked 20 November 2026. This is a 2027 planning guide, not a certification statement or legal advice.