Hub4AI company logo

Hub4AI Subprocessor and Engine Register

Last updated: 5 September 2026

1) Purpose and scope

This register identifies the subprocessors and AI engine providers that Hub4AI may use to provide AI4Business and related services. It is intended to support transparency for Customers, including the information required by the applicable data processing agreement, privacy notice and customer contract.

The current entries will state, where applicable, the provider legal entity, service or processing purpose, relevant product or feature, typical processing location, international transfer mechanism, retention information and any applicable customer-choice or regional-routing condition.

2) How to use this register

A provider is included only when it is used or may be used for a live service route, an enabled feature or an approved operational purpose. An AI engine is used only when selected or permitted through the relevant service configuration. A Customer may have controls over the engines, regions or connectors available in its own environment, as described in the applicable contract and product documentation.

This register is not a complete description of every third-party service that a Customer may independently connect to or use. Customer-enabled connectors and destination systems are governed by the Customer’s own instructions and the relevant third-party terms.

3) Register entries

The current infrastructure and AI provider entries are maintained below: 

a) Microsoft Azure (Microsoft Foundry) + OpenAI GPT

  • What is stored: Prompts, completions, embeddings, files, abuse samples; Prompts and completions, embeddings, uploaded/fine-tuning files, stateful data (Responses API, Assistants threads, stored completions), plus a sampled copy of flagged prompts/completions kept in a per-resource abuse-monitoring store.
  • How stored data may be used: Service delivery + abuse detection only; Used to deliver the service and to detect uses that violate product terms. Flagged samples are reviewed automatically by AI models by default, with human review by authorized Microsoft employees where necessary (EEA staff for EEA deployments).
  • How long data is kept: ~30 days for abuse logs, indefinite for what you store; Abuse-monitoring store retained up to ~30 days per current support guidance (the public docs no longer print an exact number). Customer-controlled data — files, threads, fine-tunes — persists until you delete it, encrypted at rest in your chosen Azure geography.
  • Data used for future training: No; Not used by the model provider to improve their models, and not used to train foundation models without your explicit instruction.
  • Data used for product development: No; Customer Data, Prompts, and Completions are not used to improve Microsoft or third-party products or services without your explicit permission or instruction.
  • User prompts used for product development: No; Same clause covers prompt text specifically — no separate carve-out exists.
  • Max days data is stored: 30 days typical ceiling, unlimited if self-stored; The abuse-monitoring ceiling is ~30 days; anything you choose to persist (files, fine-tunes, threads) has no automatic expiry.
  • Customer can opt out of storage: Yes, but gated; Modified Abuse Monitoring removes storage and human review entirely, but requires an approved application, is limited to managed/EA-MCA customers, and takes about 5–10 business days to process.
  • Compatibility with AI Act: Full Code of Practice signatory, strong tooling; Microsoft and OpenAI are both full signatories across all three CoP chapters, giving a presumption of GPAI conformity. Azure adds EU Data Boundary residency and ISO/IEC 42001 certification that support your own Art. 12/26 deployer obligations — but registration, risk management, and human oversight design remain yours.
  • Terms source: https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy · https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/abuse-monitoring

b) Microsoft Azure (Microsoft Foundry) + Anthropic Claude

  • What is stored: Prompts and outputs only; No separate embeddings/fine-tuning store — just conversational inputs and outputs, processed either on Azure infrastructure or on Anthropic’s, depending on the hosting option selected.
  • How stored data may be used: Anthropic reviews flagged content, not Microsoft; Anthropic is the seller, operator, and independent data processor for this traffic — not Microsoft. Automatic safeguards may flag content for Anthropic Trust & Safety review on an exceptions-only basis.
  • How long data is kept: 30 days typical, up to 7 years worst case; ~30 days default deletion of inputs/outputs; up to 2 years for content flagged by trust & safety systems; up to 7 years for the resulting classification scores. Covered (Mythos-class) models force a mandatory 30-day retention on every platform, no exceptions.
  • Data used for future training: No; Anthropic’s commercial terms bar training on this traffic without your express permission.
  • Data used for product development: No; Anthropic’s opt-in Development Partner Program is a first-party-API-only offering and is not extended to third-party platforms like Azure.
  • User prompts used for product development: No; Same restriction — no separate treatment for prompt text.
  • Max days data is stored: 7 years worst case (classifier scores); Typical case is 30 days; the 7-year figure only applies to derived classification scores from flagged content, not raw conversation text (that caps at 2 years).
  • Customer can opt out of storage: Partially, and not through Microsoft; There’s no Microsoft-side toggle — any zero-retention arrangement has to be negotiated directly with Anthropic, and it doesn’t cover Covered Models or already-flagged content.
  • Compatibility with AI Act: Full CoP signatory, but split accountability; Anthropic is a full Code of Practice signatory. The catch is structural: because Anthropic — not Microsoft — is the data processor here, Microsoft’s ISO 42001 scope, EU Data Boundary, and Product Terms don’t extend to this model traffic, so your Art. 26 evidence chain splits across two vendors.
  • Terms source: https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/claude-models/data-privacy · https://www.anthropic.com/legal/data-processing-addendum

c) Microsoft Azure (Microsoft Foundry) + xAI Grok

  • What is stored: Same surfaces as Azure OpenAI; Grok is sold, hosted, billed, and operated by Azure directly (“Models sold by Azure”), so it shares the same prompts/completions/abuse-sample storage pattern as the Azure-OpenAI row.
  • How stored data may be used: Service delivery + abuse detection, xAI has no access; Identical commitment to the Azure-OpenAI row — the model provider itself never sees your traffic.
  • How long data is kept: ~30 days abuse store, unlimited self-stored; Same structure as Azure OpenAI, with one caveat: several Grok versions are Preview, and Azure preview features may follow different privacy practices, including for abuse monitoring.
  • Data used for future training: No; Covered by the same “Models sold by Azure” no-training commitment.
  • Data used for product development: No; Same explicit-permission requirement as the Azure-OpenAI row.
  • User prompts used for product development: No; Same.
  • Max days data is stored: 30 days typical ceiling; Same as Azure OpenAI, with the added uncertainty that Preview-labeled Grok versions may not follow the stated policy exactly.
  • Customer can opt out of storage: Yes, but gated; Same Modified Abuse Monitoring path and eligibility requirements as Azure OpenAI.
  • Compatibility with AI Act: Partial — weakest row in the table; xAI signed only the Safety and Security chapter of the Code of Practice, not the Transparency or Copyright chapters, meaning it must demonstrate those obligations through alternative adequate means rather than through the presumption of conformity. Combined with several Grok versions being Preview, this is the highest-uncertainty row for any Annex III use case.
  • Terms source: https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy · https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure

d) Google Cloud (Vertex AI) + Gemini

  • What is stored: Cached I/O plus flagged prompts; Cached inputs/outputs (project-level, can be disabled), prompts logged only when safety classifiers flag suspicious activity, plus any grounding, tuning, or RAG data you supply.
  • How stored data may be used: AUP enforcement only, never training; Logged prompts are used solely to examine whether the AUP or Prohibited Use Policy was violated. Authorized Google employees may review flagged content. A Section 17 Training Restriction covers all managed models on Vertex.
  • How long data is kept: 24h cache, 90 days for flagged logs, 30 for grounding; Cache expires within 24 hours in the serving data center; flagged-prompt logs are kept up to 90 days in your selected region, not CMEK-encrypted; grounding data (Search/Maps) is kept 30 days with no way to disable it.
  • Data used for future training: No; The Training Restriction explicitly bars using this data to train or fine-tune any AI/ML model.
  • Data used for product development: Partial; Abuse logs are not used for product development, but grounding-with-Search/Maps data may be used for debugging and testing of the systems that support grounding — a use you cannot switch off if you use the feature.
  • User prompts used for product development: Partial; Same exception as above — it’s the grounding pathway specifically, not general prompt text, that carries this use.
  • Max days data is stored: 90 days for flagged logs; The 90-day figure is the binding ceiling for anything logged; cache and grounding data are shorter but structurally different (grounding can’t be turned off).
  • Customer can opt out of storage: Partial; Abuse logging can be waived by request (customers under a Google Cloud Master Agreement are exempt by default) and caching can be disabled per project — but grounding storage cannot be disabled if the feature is used.
  • Compatibility with AI Act: Full CoP signatory, strong tooling; Google is a full Code of Practice signatory. Regional/EU endpoints, Access Transparency, VPC Service Controls, and Cloud Audit Logs support your Art. 12/15 obligations — but use regional (not global) endpoints if data residency is part of your compliance argument.
  • Terms source: https://cloud.google.com/vertex-ai/generative-ai/docs/vertex-ai-zero-data-retention · https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/abuse-monitoring

e) Google Cloud (Vertex AI Model Garden) + Anthropic Claude

  • What is stored: Same surfaces as Gemini, Google is the processor; Requests route through Vertex AI endpoints; Google — not Anthropic — is confirmed as the data processor for this combination.
  • How stored data may be used: AUP enforcement; Advanced AI models get extra scrutiny; Standard use mirrors the Gemini row. Models designated “Advanced AI” (Claude Mythos and Fable, all versions; Opus ≥4.7 and Sonnet ≥5 when advanced_ai_enabled is true) are subject to a more rigorous monitoring system with automated investigative tooling.
  • How long data is kept: 90 days standard, 30 days for Advanced AI logging; Standard: 24h cache + 90-day flagged logs, same as Gemini. Advanced AI models: ALL prompts and responses are logged and stored for up to 30 days, not CMEK-encrypted, regardless of flagging.
  • Data used for future training: No; Covered by the same Vertex Training Restriction as Gemini.
  • Data used for product development: No; Unlike Gemini’s grounding exception, logged Claude data on Vertex is not used to train or fine-tune models — no carve-out applies here.
  • User prompts used for product development: No; Same.
  • Max days data is stored: 90 days standard, 30 days mandatory for Advanced AI; The Advanced AI figure is lower but unconditional — every prompt and response is captured, not just flagged ones.
  • Customer can opt out of storage: Partial; Standard models follow the same opt-out mechanisms as Gemini. For Advanced AI models, opting out may not be possible at all — confirm with your account team before relying on it.
  • Compatibility with AI Act: Full CoP signatory, single accountability line; Anthropic is a full CoP signatory, and because Google (not Anthropic) is the processor here, you avoid the two-vendor split seen in the Azure+Claude row. Watch GDPR storage-limitation exposure on Advanced AI models’ mandatory 30-day full logging.
  • Terms source: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-partner-models · https://cloud.google.com/terms/advanced-ai-safety-addendum

f) AWS (Amazon Bedrock) + Anthropic Claude

  • What is stored: By default, nothing; specific models require it; Bedrock uses a zero data retention (ZDR) posture by default — Amazon Bedrock doesn’t store model inputs or outputs. Certain models override this: Claude Fable 5 and Claude Mythos 5 require provider_data_share mode, under which prompts and completions are shared with Anthropic and retained.
  • How stored data may be used: Abuse detection and, for Covered Models, Anthropic trust & safety review; For models under provider_data_share, data is used for abuse detection and potential human review by Anthropic, as required by Anthropic for access to these models. For other models under default mode, AWS may retain classifier-flagged traffic for its own abuse detection, without sharing it with Anthropic.
  • How long data is kept: Up to 30 days for the models that require it; For Claude Fable 5/Mythos 5, inputs and outputs are retained for up to 30 days and shared with Anthropic. If cross-region inference is enabled, retained data is stored in the destination region, not necessarily your origin region.
  • Data used for future training: No; Anthropic maintains a strict policy of never training on customer data delivered via Bedrock, and the retained-for-abuse-detection data is no exception.
  • Data used for product development: No; No product-development use is documented for retained Bedrock traffic — retention is scoped strictly to safety and abuse-prevention purposes.
  • User prompts used for product development: No; Same.
  • Max days data is stored: 30 days, only for the models that mandate it; Models with allowed_modes including none are never retained at all; only Fable 5/Mythos 5 (and any future model AWS marks the same way) hit the 30-day ceiling.
  • Customer can opt out of storage: Yes for most models by default; gated for Covered Models; Default posture is already zero retention for the majority of Claude models on Bedrock. For the models that require provider_data_share, full ZDR can be requested but is evaluated per-account, per-model in coordination with Anthropic, and is not guaranteed. Org-wide ZDR can also be enforced via AWS Service Control Policies.
  • Compatibility with AI Act: Full CoP signatory, AWS provides the audit surface; Anthropic is a full Code of Practice signatory; Amazon separately appears on the EC signatory list in its own capacity as a GPAI provider, which does not substitute for Anthropic’s own obligations. Data resides in AWS-selected regions (watch cross-region inference for residency claims); AWS’s default ZDR posture is unusually strong evidence toward data-minimisation arguments under Art. 10.
  • Terms source: https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html · https://docs.aws.amazon.com/bedrock/latest/userguide/abuse-detection.html · https://platform.claude.com/docs/en/manage-claude/api-and-data-retention

g) AWS (Amazon Bedrock) + OpenAI GPT

  • What is stored: Zero by default; specific GPT models retain flagged traffic; Same Bedrock-wide ZDR default. For specific OpenAI GPT models (GPT-5.4, GPT-5.5, GPT-5.6 family, Daybreak Red/Blue), AWS retains classifier-flagged traffic for up to 30 days for automated offline abuse detection. Responses API requests default to store: true, so response data (including its input and output) is retained for 30 days regardless of flagging, to support statefulness.
  • How stored data may be used: Automated abuse detection only, processed programmatically; Flagged traffic is processed programmatically for offline abuse detection — not shared with OpenAI, and neither Daybreak model requires you to opt in to sharing data with OpenAI. Inference data across the board is not used for model training.
  • How long data is kept: 30 days for flagged/stateful traffic; Classifier-flagged traffic: up to 30 days. Responses API stateful data: 30 days by default when store: true. Chat Completions API: zero data retention — no conversation data stored between requests at all.
  • Data used for future training: No; Your inference data is not used for model training, stated explicitly even for the specialized Daybreak cyber-defense models.
  • Data used for product development: No; No product-development use is documented; retention is scoped to abuse detection and statefulness support only.
  • User prompts used for product development: No; Same.
  • Max days data is stored: 30 days; Applies uniformly to both the classifier-flagged abuse path and the Responses API stateful-storage path.
  • Customer can opt out of storage: Yes, but Responses API default needs an explicit change; Chat Completions API is zero-retention by default. Responses API stores by default (store: true) — you must explicitly set store: false, though AWS notes this alone does not guarantee full ZDR; for models that mandate retention, request ZDR through your AWS account team, evaluated per-account and per-model.
  • Compatibility with AI Act: Full CoP signatory; open-weight variants need individual review; OpenAI is a full Code of Practice signatory for its hosted GPT models. Note that AWS also separately offers OpenAI’s open-weight GPT-OSS models via Bedrock Custom Model Import — these are self-hosted artifacts under Apache 2.0 licensing rather than OpenAI-operated endpoints, so the CoP presumption and OpenAI’s own data commitments don’t automatically extend to that deployment path; treat open-weight imports as a separate compliance review.
  • Terms source: https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html · https://docs.aws.amazon.com/bedrock/latest/userguide/abuse-detection.html · https://developers.openai.com/api/docs/guides/amazon-bedrock

h) OpenAI (direct) + GPT via API

  • What is stored: Abuse logs, application state, uploaded/fine-tuning files; Abuse-monitoring logs that may contain prompts and responses plus derived metadata such as classifier outputs, generated for all API feature usage. Also: application state for stateful features, uploaded files, fine-tuning data, vector stores.
  • How stored data may be used: Enforcing usage policies and providing the service; Retention exists to enforce OpenAI’s Usage Policies and agreements and to mitigate harmful uses, and to provide the services themselves.
  • How long data is kept: 30 days for logs, indefinite for stored objects; Up to 30 days for API inputs/outputs, then removed unless legally required to retain. Responses API stores response data for at least 30 days when store is true. Fine-tuning files persist until you delete them. Assistants objects not deleted via API or dashboard are retained indefinitely.
  • Data used for future training: No; Not used for model training by default on API/business tiers.
  • Data used for product development: No, as stated; No explicit product-development use is claimed beyond service provision — though the “to provide the services” language is broader than Microsoft’s explicit carve-out and is worth clarifying contractually.
  • User prompts used for product development: No, as stated; Same caveat applies to prompt text specifically.
  • Max days data is stored: 30 days typical, indefinite for un-deleted stored objects; The 30-day figure covers the abuse log; anything you leave in Assistants threads or vector stores has no automatic expiry.
  • Customer can opt out of storage: Yes, but gated; Zero Data Retention or Modified Abuse Monitoring require prior approval, a qualifying use case, an enterprise API agreement, and are limited to eligible endpoints.
  • Compatibility with AI Act: Full CoP signatory, thinnest residency story; OpenAI is a full Code of Practice signatory. Unlike the cloud-hosted rows, there’s no EU Data Boundary equivalent or native cloud audit logging — you build your own Art. 12 record-keeping from scratch.
  • Terms source: https://developers.openai.com/api/docs/guides/your-data · https://openai.com/enterprise-privacy/

i) Anthropic (direct) + Claude via API

  • What is stored: Nothing by default; specific stateful features do; Conversation content is not retained by default — only what’s technically necessary for a feature to work, purged on the shortest practical TTL. Exceptions: batch processing (29 days), code-execution container data (up to 30 days), Files (until deleted/expired), Managed Agents transcripts (until you delete them).
  • How stored data may be used: Service delivery, feature function, and policy enforcement; Retained data supports feature functionality and Usage Policy enforcement, and is never used for model training without your express permission.
  • How long data is kept: ~30 days typical, up to 7 years worst case; Default commercial policy deletes inputs/outputs within ~30 days. Content flagged by automated trust & safety systems may be retained up to 2 years — even under ZDR or HIPAA arrangements. Classification scores derived from that review persist up to 7 years.
  • Data used for future training: No; Retained data is never used for training without your express permission; an opt-in Development Partner Program exists for organizations that choose to contribute data.
  • Data used for product development: No by default; opt-in only; Same Development Partner Program is the only channel, and it requires explicit enrollment.
  • User prompts used for product development: No by default; opt-in only; Same.
  • Max days data is stored: 7 years worst case (classifier scores), 30 typical, 0 under ZDR; The ceiling is structural — even the strictest ZDR/HIPAA arrangement cannot suppress the 2-year flagged-content and 7-year classifier-score retention.
  • Customer can opt out of storage: Yes, but gated; Zero Data Retention is available via sales approval, enabled per organization (doesn’t extend automatically to sibling orgs on the same account). Excluded from ZDR: Console usage, Managed Agents, consumer plans, Files, batches, code execution, and Claude Fable 5/Mythos 5, which mandate 30-day retention.
  • Compatibility with AI Act: Full CoP signatory, defensible documentation; Anthropic is a full Code of Practice signatory and publishes model system cards that map well onto Art. 53 documentation duties. The 7-year classifier-score retention is defensible under the AI Act itself but sits awkwardly against GDPR’s storage-limitation principle — worth flagging in any joint AI Act/GDPR review.
  • Terms source: https://platform.claude.com/docs/en/manage-claude/api-and-data-retention · https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data

4) Maintenance and updates

Hub4AI may update this register from time to time to reflect a new or replacement subprocessor, an enabled or disabled AI engine, a changed processing location, or another material change to the service supply chain. The date above records the most recent update to this version of the register.

Where the applicable data processing agreement requires advance notice of a new or replacement subprocessor, Hub4AI will provide that notice and apply the related Customer objection process. In the event of a conflict, the applicable Order, data processing agreement and mandatory law prevail over this register.

Questions about this register may be sent to support@hub4ai.com.

Exit mobile version