The AI Vendor Telemetry Audit

Confidence Tiers Used in This Report

Every key finding in this report carries a confidence tier:

• [VERIFIED] — Confirmed by vendor documentation, regulatory finding, or direct technical analysis. Source cited.
• [INFERRED] — Logical conclusion from verified facts, but not explicitly stated by the vendor. Reasoning shown.
• [UNKNOWN] — Vendor documentation is silent or contradictory. Gap flagged for client-specific audit.
• [REPORTED] — Reported by reputable press or secondary source; not independently confirmed against primary documentation. Source and date cited.

Executive Summary

Enterprise adoption of cloud-hosted generative AI has outpaced legal and security review. Organizations are feeding proprietary data, customer records, and strategic queries into APIs owned by three vendors — Microsoft, Google, and Amazon Web Services — without a clear inventory of what those vendors log, retain, analyze, or disclose.

This audit examines the telemetry and data-handling practices of AI vendors across two tiers. First, the three dominant hyperscaler platforms that resell foundation models inside their own governance wrappers:

• Microsoft Azure OpenAI / Foundry
• Google Cloud Vertex AI / Gemini
• Amazon Web Services Bedrock / SageMaker

Second, the direct model-vendor APIs many teams call without a hyperscaler intermediary — where the default data posture diverges most sharply:

• Anthropic (Claude) — no-train by default, ZDR and BAA available
• OpenAI (direct API) — no-train by default, ZDR on eligible endpoints
• Moonshot AI (Kimi) — trains by default, no BAA, prepaid-balance seizure clause

Key finding: [VERIFIED] While all three vendors promise that customer prompts are not used to train foundation models, each platform generates a substantial metadata footprint — API call logs, token counts, latency metrics, IP addresses, identity objects, and abuse-monitoring telemetry — that creates compliance surface area most enterprises have not assessed. The May 2026 joint PIPEDA investigation into OpenAI [VERIFIED — OPC Findings #2026-002] confirms that Canadian regulators are now actively scrutinizing these practices.

September 2026 update [VERIFIED]: In a September 8, 2026 public statement about its models' use of product-usage data, OpenAI said it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." It is a rare vendor-authored acknowledgment that the usage-derived de-identification channel may feed model improvement — a channel the training commitments audited in this report do not cover. Section 5B places that admission, and September's cross-border prompt-routing and open-weight migration reporting, in this audit's framework.

For organizations subject to PIPEDA, provincial privacy laws, GDPR, or sector-specific regulations (healthcare, finance, defense), the default configuration of cloud AI APIs presents material risk. This report documents those risks and outlines the audit framework elect-rix applies when assessing client AI infrastructure.

1. The Stakes: Why Telemetry Matters

When an employee submits a prompt to a cloud AI model, the organization leaks more than just the text of the query. The transaction produces:

The misconception: "We have an enterprise agreement that says our data is not used for training."

The reality: Training-data commitments address only one vector. They do not eliminate logging, do not prevent subpoena-driven disclosure, and do not exempt the vendor from retaining prompts for abuse monitoring, debugging, or legal compliance.

2. Methodology

This audit is based on:

1. Published vendor documentation (Microsoft Learn, AWS Documentation, Google Cloud Terms) as of June 2026 — September 2026 addendum material in Section 5B is sourced and dated inline
2. Regulatory findings — Office of the Privacy Commissioner of Canada, PIPEDA Findings #2026-002 (May 6, 2026)
3. Independent technical analysis of API logging behavior and diagnostic telemetry
4. Enterprise architecture reviews conducted by elect-rix in production environments

This report does not rely on vendor marketing materials. Where documentation is ambiguous, we note the gap and classify it as an unverified risk.

Scope

This audit covers the default and commonly-deployed configurations of three hyperscaler AI platforms as documented and observed through June 2026: Microsoft Azure OpenAI / Foundry, Google Cloud Vertex AI / Gemini, and Amazon Web Services Bedrock / SageMaker. It addresses telemetry, logging, retention, and metadata exposure relevant to Canadian data-sovereignty and PIPEDA obligations.

Limitations

• Vendor data-handling practices change frequently; findings reflect documentation and observed behaviour current as of the publication date below and may not reflect subsequent changes.
• Enterprise agreements, regional deployments, and private-networking configurations can materially alter telemetry behaviour. Organization-specific exposure requires a direct assessment.
• Where vendor documentation is silent or ambiguous, items are classified as unverified risks rather than confirmed practices.
• This report is informational and is not legal advice. It does not establish an attorney-client relationship.

Published: June 16, 2026  ·  Version: 1.2 (updated September 23, 2026)  ·  Next scheduled review: December 2026

3. Microsoft Azure OpenAI / Foundry

3.1 Official Data Commitments

Microsoft publishes five core commitments for Azure OpenAI and Foundry Models:

• Prompts and completions are NOT available to other customers
• They are NOT available to OpenAI or other model providers
• They are NOT used to improve models without explicit permission
• They are NOT used to train generative AI foundation models without permission
• Fine-tuned models are exclusively available to the creating customer

Source: Microsoft Learn — Data, Privacy, and Security for Models Sold by Azure

3.2 What Microsoft Actually Processes

Microsoft processes five categories of data:

1. Prompts and generated content — inputs and outputs
2. Uploaded data — files for fine-tuning, assistants, batch processing
3. Stateful entity data — message history, thread content, stored completions
4. Augmented data — retrieved documents for RAG/grounding
5. Training & validation data — prompt-completion pairs for custom models

3.3 Abuse Monitoring — The Hidden Retention Window

Microsoft acknowledges that prompts and completions are retained for up to 30 days for abuse detection and mitigation. During this window:

• Content is analyzed for harmful output, jailbreak attempts, and code-of-conduct violations
• Human reviewers may review flagged content
• The data resides within the customer-specified geography but may move between regions within that geography for operational purposes

Critical caveat: "Global" and "DataZone" deployment types expand processing geography significantly. A "Global" deployment may process data in any region where the model is deployed, though data at rest remains in the designated geography.

3.4 Native Logging Gaps — What Enterprises Cannot See

Azure OpenAI emits three log types via diagnostic settings:

As of January 2024: Entra ID object ID was added to RequestResponse logs. Microsoft now logs who made the call but not what was in it.

Enterprise impact: Native logs are insufficient for chargeback, input/output auditing, and token-level forensic analysis. Organizations must deploy an API gateway (Azure API Management or equivalent) to capture the telemetry they actually need — which means the gateway layer now becomes a second point of data exposure.

3.5 Subpoena and Disclosure Risk

Microsoft operates under U.S. law. The Microsoft Products and Services Data Protection Addendum (DPA) governs data handling, but like all U.S.-based cloud providers, Microsoft is subject to:

• U.S. CLOUD Act — compels disclosure of data held by U.S. providers regardless of physical location
• FISA 702 / PRISM — upstream collection of communications content
• National Security Letters — gag-ordered data requests

Microsoft publishes a transparency report but cannot disclose the volume of classified requests.

4. Google Cloud — Vertex AI & Gemini

4.1 Data Use Framework

Google Cloud's position on AI data usage is tiered:

Critical distinction: Google's consumer and enterprise AI stacks are not the same product. An employee using gemini.google.com with a personal account exposes corporate data to consumer-grade terms. Shadow AI — employees using consumer tools for work — is the fastest-growing ungoverned data channel in enterprise IT.

4.2 The Google AI Studio Trap

Google AI Studio (formerly MakerSuite) operates under consumer terms unless the organization activates a Google Cloud billing account. Many developers begin prototyping in AI Studio and later "migrate" to Vertex AI without understanding that the prototyping phase subjected their data to consumer data-use policies.

4.3 Telemetry and Logging

Google Cloud logs API activity through Cloud Audit Logs. For Vertex AI:

• Admin Activity logs — record API calls that modify resources (model deployment, endpoint creation)
• Data Access logs — record API calls that read data (must be explicitly enabled; disabled by default due to cost/volume)
• System Event logs — record Google-initiated actions

What is NOT logged natively: Prompt content and response content are not captured in Cloud Audit Logs. However:

• Latency, token count, and error rates are exposed via Cloud Monitoring
• Request/response logging requires customer-implemented instrumentation
• Model Garden (third-party models hosted on Vertex AI) may have separate telemetry terms

4.4 Data Residency and the Multi-Region Problem

Google Cloud allows data location selection for specific services. However:

"Google may replicate Customer Data within other Regions in the same country (or countries of the Multi-Region) for backup, reliability, debugging, support, maintenance, or security."

Source: Google Cloud Service Specific Terms

This replication authority means that a "Canada" region selection does not guarantee data never touches U.S. infrastructure for operational purposes. Debugging and support activities are explicitly exempted from data-location commitments.

4.5 Pre-GA Offerings — The Liability Void

Google's Pre-GA terms (Alpha, Beta, Preview, Experimental) contain explicit disclaimers:

"PRE-GA OFFERINGS ARE PROVIDED 'AS IS' WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES... [They] are not covered by any SLA or Google indemnity."

Enterprise AI teams frequently adopt preview features (e.g., early Gemini model versions, agent frameworks) without legal review of these terms. The result: production workloads running on software with no warranty, no SLA, and no data-location guarantees.

5. Amazon Web Services — Bedrock & SageMaker

5.1 The Shared Responsibility Model

AWS applies its standard shared responsibility model to AI services:

• AWS responsibility: Global infrastructure security
• Customer responsibility: Content security, access configuration, encryption management

Source: AWS Bedrock Data Protection Documentation

5.2 Bedrock Data Handling

AWS Bedrock uses a Model Deployment Account architecture:

• One deployment account per model provider per region
• Accounts are owned by the Bedrock service team — not by the model provider (Anthropic, Meta, etc.)
• Model providers have no access to these accounts after model delivery
• Providers cannot access Bedrock logs or customer prompts/completions

AWS commitments:

• Customer content is not used for marketing or advertising
• Content is not disclosed without customer agreement except under legal compulsion
• Customers choose the AWS Region(s) for storage
• AWS attempts to redirect governmental requests to the customer

5.3 The CloudTrail Blind Spot

AWS CloudTrail captures API calls for Bedrock and SageMaker. However:

• CloudTrail logs the API action (e.g., `InvokeModel`) but not the payload
• Prompt and response content are not written to CloudTrail
• Token usage is available via CloudWatch metrics but requires explicit metric stream configuration
• VPC Flow Logs capture network metadata but not application-layer content

Enterprise impact: Like Azure, AWS requires an intermediary proxy or application-layer logging to audit inputs and outputs. This proxy becomes a critical security control — and a critical vulnerability if misconfigured.

5.4 Third-Party Model Risk

Bedrock hosts models from Anthropic, Meta, Stability AI, Cohere, and others. While AWS states that model providers cannot access customer data, the legal chain is complex:

1. Anthropic trains Claude on undisclosed datasets
2. AWS deploys a copy of Claude in its Model Deployment Account
3. Customer prompts go to the AWS-hosted copy
4. AWS does not train on customer data

The unverified risk: Whether Anthropic's training data included legally problematic sources (copyrighted works, personal data scraped without consent) affects the compliance posture of every output generated by the model. AWS does not warrant the provenance of third-party model training data.

5A. Direct Model-Vendor APIs: Anthropic, OpenAI, Moonshot

The hyperscaler platforms above resell foundation models inside their own governance wrappers. But many organizations call the model vendors directly — api.anthropic.com, api.openai.com, api.moonshot.ai. The default data posture differs sharply between these vendors, and the gap is the single most consequential procurement decision most teams never consciously make.

Anthropic (Claude) — No-Train by Default

• Training: Anthropic does not train its models on API inputs or outputs by default under the Commercial Terms of Service. No opt-out toggle is required — it is the standing default.
• Zero Data Retention (ZDR): Available on request for qualified accounts. With ZDR, inputs and outputs are not retained after the request completes.
• HIPAA: Anthropic will execute a Business Associate Agreement (BAA), making a compliant PHI path possible on the API.

Verdict [VERIFIED — Anthropic Commercial Terms]: Of the direct model vendors, Anthropic carries the most privacy-favorable defaults — no-train standing, ZDR available, HIPAA path open.

OpenAI (Direct API) — No-Train by Default, 30-Day Abuse Window

• Training: API data submitted after March 1, 2023 is not used for training by default.
• Retention: Inputs and outputs retained up to 30 days for abuse monitoring, then deleted unless legally required. Zero Data Retention available for eligible endpoints on approved accounts.
• HIPAA: BAA available on the API Platform.

Verdict [VERIFIED — OpenAI Enterprise Privacy, Jan 2026]: Strong defaults, near-parity with Anthropic, with the practical difference that ZDR is endpoint-gated and the standard path keeps a 30-day window.

Moonshot AI (Kimi) — Trains by Default

• Training: Per the Moonshot Terms of Service (last updated May 27, 2026), Content — defined as both your prompts and the model's outputs — may be used to "provide, maintain, develop, support, and improve the Services," which expressly includes model development, unless otherwise agreed in writing. This is training-by-default.
• Opt-out: No self-serve toggle. An exclusion path exists only via a bespoke enterprise arrangement — negotiate-and-sign, not a setting.
• HIPAA / PHI: Prohibited outright. PHI may not be submitted to the API at all. There is no BAA path.
• Enforcement: On a violation, and without prior notice, Moonshot may suspend or close the account and "freeze and confiscate the recharged amount" — i.e. seize prepaid balance.

Verdict [VERIFIED — Moonshot TOS, May 27, 2026]: Moonshot's defaults are the inverse of Anthropic's. For any workload touching regulated, personal, or proprietary data, the direct Moonshot API is the highest-risk option surveyed in this report.

The procurement lesson: "It's an API, they're all the same" is false. The vendor you pick is your privacy policy. For organizations that cannot accept any of these postures, Section 9's local-deployment path removes the question entirely.

5B. September 2026 Addendum: The Vendors' Own Words

The vendor commitments documented above were written by the vendors. September 2026 supplied the test.

5B.1 OpenAI: "We Cannot Rule Out" [VERIFIED]

On September 8, 2026, OpenAI announced its models had produced a claimed solution to one instance of the Navier-Stokes existence and smoothness problem — a Clay Mathematics Institute Millennium Prize problem. The claim immediately became a priority dispute: mathematician Tristan Buckmaster (NYU) stated, roughly 12 hours before the announcement, that he and Levent Alpöge had been working on closely related problems inside OpenAI's own coding agent (Codex) and had been tipped that their "progress had been passed to OpenAI" (BBC News, Sept 8, 2026).

OpenAI denied direct use of their work but stated, in its own words:

"While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
— OpenAI statement, September 8, 2026 (as published by BBC News)

[VERIFIED — OpenAI public statement, Sept 8, 2026] This is a rare vendor-authored acknowledgment that the "de-identified data derived from usage" channel exists and may feed model improvement. It is a materially different channel from the prompt-and-completion commitments audited in Section 5A:

• The API no-train commitment covers content submitted to eligible endpoints. The admission concerns derivative usage data — a category most enterprise compliance reviews do not separately assess.
• A vendor that "cannot rule out" its own model-improvement pipeline cannot certify to a customer that confidential work conducted inside its tools stayed out of that pipeline.
• The exposure was surfaced by the researchers, not disclosed by the vendor. Nature's commentary (Sept 17, 2026) draws the general lesson: researchers can be "inadvertently sharing — and absorbing — ideas through chatbots."

Audit implication [INFERRED]: Enterprise AI governance that stops at "our prompts are not used for training" is incomplete. Usage-derived de-identified data must be assessed as its own data flow, and for confidential-development workloads the only configuration that closes the channel is local deployment (Section 9).

5B.2 Prompt Routing Is a Supply Chain [REPORTED]

On September 22, 2026, The Information reported that China's Cyberspace Administration is investigating DeepSeek and Moonshot AI after Anthropic alleged that both had routed sensitive user data — including police case-management queries and camera footage — to Claude models without user awareness. The report cites people with knowledge of the matter; the parties have not confirmed the details.

Sovereignty implication [REPORTED — The Information, Sept 22, 2026]: Where a prompt travels is decided by a chain of vendors — your vendor's vendor — and the September reporting shows that chain operating across borders without user knowledge. Data-sovereignty exposure is not a Canada-to-US problem; it is a routing-transparency problem that runs in every direction.

5B.3 The Economics Arrived: Open-Weight Migration [REPORTED]

On September 21, 2026, Bloomberg reported that Harvey, a legal-AI company, moved its flagship product off frontier-lab APIs to open-weight models after agentic usage collapsed its gross margin from roughly +50% to -50% between January and June 2026; Abridge, Decagon, Ramp, Rogo, and Canva are making similar shifts. The margin figures are single-sourced (a person familiar with the matter), but the companies' public model choices corroborate the direction.

Market implication [INFERRED]: The local/open-weight alternative in Section 9 once argued from compliance alone. It now also argues from unit economics: agentic workloads multiply token consumption, and per-token API pricing converts that multiplication directly into negative margin. For cost-sensitive regulated buyers, the open-weight path carries a business case that does not depend on privacy arguments at all.

6. The Metadata Layer: What Vendors Don't Talk About

Even when prompt content is protected, the metadata surrounding AI API calls reveals significant organizational intelligence:

This metadata is not covered by "your data is not used for training" promises. It is operational telemetry that vendors retain for billing, capacity planning, product improvement, and abuse prevention.

The strategic risk: A competitor or adversary with access to this metadata (via breach, subpoena, or insider action) could reconstruct organizational priorities without ever seeing a single prompt.

7. Regulatory Context: The Canadian Crackdown

7.1 PIPEDA Findings #2026-002 — OpenAI Investigation

On May 6, 2026, the Office of the Privacy Commissioner of Canada, joined by Quebec, British Columbia, and Alberta privacy authorities, issued findings against OpenAI OpCo, LLC regarding ChatGPT.

Key findings:

• Jurisdiction affirmed: Canadian privacy laws apply to OpenAI despite its U.S. corporate structure. "The Internet has no borders."
• Overbroad collection: OpenAI's collection of publicly accessible and licensed data for training was found inappropriate — not merely incidental, but systematic and excessive
• Consent deficiencies: OpenAI failed to obtain valid consent for collection, use, and disclosure of personal information
• Transparency failures: Users were not adequately informed about what data was collected and how it was used
• Accuracy risks: Generated personal information (hallucinations about individuals) creates defamation and misrepresentation risks

Source: Office of the Privacy Commissioner of Canada — PIPEDA Findings #2026-002

7.2 Implications for Enterprise Cloud AI Users

The OpenAI investigation has cascading implications:

1. Liability chain: If OpenAI's training data collection was non-compliant, outputs generated from that training may carry tainted provenance. Organizations using these outputs for customer-facing applications assume downstream liability.
2. Shadow AI exposure: Employees using consumer ChatGPT (not enterprise Azure OpenAI) subject the organization to the exact practices found non-compliant by the OPC.
3. Due diligence standard: The investigation establishes that "we have an enterprise agreement" is insufficient. Organizations must now verify:
4. What training data was used for the specific model version they deploy
5. Whether that training data included Canadian personal information collected without consent
6. Whether outputs can be traced to non-compliant training sources

7.3 GDPR and Cross-Border Transfer Risk

For organizations with EU operations or EU data subjects:

• The Schrems II decision invalidated the Privacy Shield framework
• Standard Contractual Clauses (SCCs) remain valid but require Transfer Impact Assessments (TIAs)
• U.S. CLOUD Act and FISA 702 create structural conflicts with GDPR Article 44-49
• Meta's €1.2 billion fine (2023) for unlawful EU-U.S. data transfers set the precedent

Organizations using cloud AI for EU data subjects must document why the transfer is lawful — a burden that increases with every new model deployment.

8. Enterprise Risk Assessment Matrix

9. The Alternative: Air-Gapped & Local Deployment

The only configuration that eliminates the risks documented in this audit is local AI inference on hardware the organization owns and controls.

Trade-offs: Local deployment requires capital expenditure (hardware), technical expertise (deployment, maintenance), and capacity planning (GPU resources, model selection). For organizations where data sovereignty is non-negotiable — healthcare, defense, finance, critical infrastructure — these trade-offs are features, not bugs.

September 2026 market note [REPORTED — Bloomberg, Sept 21, 2026]: The first commercial-scale validation of this trade arrived in September 2026: Harvey moved its flagship legal-AI product to open-weight models after frontier-API agentic usage drove its gross margin negative (Section 5B.3). The local path is no longer only the compliance option; it is increasingly the margin option.

10. Recommendations

For Organizations Currently Using Cloud AI

1. Conduct a telemetry audit. Inventory every AI API call leaving your network. Document vendor, model, data classification of prompts, and applicable regulatory regime.
2. Ban consumer-tier shadow AI. Implement technical controls (DNS filtering, DLP, endpoint detection) to prevent employees from using consumer ChatGPT, Gemini, or Claude for work purposes.
3. Review deployment types. If using Azure OpenAI, avoid "Global" deployments for sensitive workloads. If using Google Vertex AI, verify data location settings and replication policies.
4. Implement gateway logging. Accept that native vendor logs are insufficient. Deploy an API gateway with input/output auditing — and secure the gateway with the same rigor as the AI service itself.
5. Document Transfer Impact Assessments. For EU or cross-border data, maintain TIAs for every cloud AI service. Update them when models change.
6. Monitor regulatory evolution. The PIPEDA OpenAI investigation is not the end. Expect sector-specific guidance from health authorities, financial regulators, and defense procurement.

For Organizations Evaluating AI Strategy

1. Classify data before selecting architecture. Not all AI workloads require the same privacy posture. Public marketing copy generation? Cloud AI may be appropriate. Patient diagnosis support? Local inference is likely mandatory.
2. Pilot with local deployment. Test model capabilities, integration patterns, and user workflows on local hardware before committing to cloud API contracts. Migration from local to cloud is trivial. Migration from cloud to local is a data-retention nightmare.
3. Negotiate contracts with telemetry carve-outs. Enterprise agreements should explicitly define what metadata vendors may retain, for how long, and for what purposes. Most standard agreements are silent on metadata.

11. What to Do Next

Book a 15-Minute AI Risk Triage — Free

Reading this report is the first step. The second step is finding out where your organization actually stands.

In a 15-minute call, elect-rix will:

1. Identify which AI services your team is using (or likely using without your knowledge)
2. Flag the highest-risk data exposure points for your industry
3. Tell you whether a full audit is warranted — or whether your current setup is reasonable

There is no obligation. No sales pitch. No follow-up sequence. If your AI usage is low-risk, we will tell you. If it is not, you will know exactly where the exposure is.

Book your free triage: contact@elect-rix.tech | elect-rix.tech

12. About This Report

This audit was prepared by elect-rix Technology Solutions, operating under RixBot Technologies Inc., a New Brunswick corporation. We design, deploy, and audit local AI infrastructure for organizations that cannot accept the data-sovereignty risks of cloud-hosted models.

Our methodology combines vendor documentation review, regulatory tracking, and hands-on infrastructure assessment. We do not sell cloud AI subscriptions. We sell control.

For a confidential assessment of your organization's AI telemetry exposure, contact elect-rix Technology Solutions via the contact page. All engagements begin with a scoped, fixed-fee assessment and a signed confidentiality agreement.

Sources and References

1. Microsoft Learn. "Data, Privacy, and Security for Models Sold by Azure in Microsoft Foundry." https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy
2. Microsoft Learn. "Implement Advanced Monitoring for Azure OpenAI in Foundry Models Through a Gateway." https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/azure-openai-gateway-monitoring
3. Journey Of The Geek. "Logging in Azure OpenAI Service." https://journeyofthegeek.com/2023/04/13/logging-in-azure-openai-service/
4. AWS Documentation. "Data Protection in Amazon Bedrock." https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html
5. AWS. "AWS Data Privacy FAQs." https://aws.amazon.com/compliance/data-privacy-faq/
6. Google Cloud. "Service Specific Terms." https://cloud.google.com/terms/service-terms
7. Google Workspace. "Generative AI Security, Compliance and Privacy." https://workspace.google.com/security/ai-privacy/
8. Office of the Privacy Commissioner of Canada. "PIPEDA Findings #2026-002: Joint Investigation of OpenAI OpCo, LLC." May 6, 2026. https://www.priv.gc.ca/en/opc-actions-and-decisions/investigations/investigations-into-businesses/2026/pipeda-2026-002/
9. Concentric AI. "Google Gemini Security Risks and Privacy Concerns Explained." https://concentric.ai/google-gemini-security-risks/
10. Anthropic. "Enterprise & Commercial Terms / Data Usage." https://www.anthropic.com/legal/commercial-terms
11. OpenAI. "Enterprise Privacy at OpenAI." (Updated January 8, 2026.) https://openai.com/enterprise-privacy/
12. OpenAI. "Data Controls in the OpenAI Platform." https://developers.openai.com/api/docs/guides/your-data
13. Moonshot AI. "Terms of Service." (Last updated May 27, 2026.) https://platform.moonshot.ai/
14. BBC News. "OpenAI says it cracked 90-year-old maths problem in 88 hours." September 8, 2026. https://www.bbc.com/news/articles/cy7zygy3rl2o
15. Nature. "Who gets credit in the AI era? OpenAI maths bombshell sparks debate." September 17, 2026. https://www.nature.com/articles/d41586-026-02910-w
16. The Information. "China Probes DeepSeek, Moonshot Over Potential Data Leaks to Anthropic." September 22, 2026. https://www.theinformation.com/articles/china-probes-deepseek-moonshot-potential-data-leaks-anthropic
17. Bloomberg. "Startups Like Harvey Embrace Open Models to Cut Reliance on Anthropic, OpenAI." September 21, 2026. https://www.bloomberg.com/news/articles/2026-09-21/startups-like-harvey-embrace-open-models-to-cut-reliance-on-anthropic-openai