Product and intended use
Define the configuration, people, information and tasks being assessed.
API org and projects
Public claim only
This profile is the API platform, not ChatGPT Business/Enterprise/Edu/consumer. Controls are organisation- and project-level (ZDR, MAM, or None). Astra for Law will appear in the API as gpt-6-astra-law; at 17 September 2026 it was initially offered via Trusted Access in ChatGPT and Codex, coming soon to the API. Harvey and Legora are named as API customers who will be able to build on it.
Ask: Which org/project, model snapshots and endpoints are in production, and is gpt-6-astra-law actually provisioned?
AU storage, not processing
Public claim only
Regional table: Australia (au.api.openai.com) Storage Yes, Processing No, requires MAM or ZDR. If the selected region does not support regional processing, OpenAI may process and temporarily store Customer Content outside the region. US and EU prefixed domains use Cloudflare Regional Services for TLS termination in-region. Checked 2026-10-10. Vendor page: “Eligible ChatGPT Enterprise, ChatGPT Edu, ChatGPT for Healthcare, and API platform customers can store sensitive customer content at rest in the U.S., Europe, UK, Japan, Canada, South Korea, Singapore, Australia, India, and the UAE to help support compliance with local data sovereignty requirements.” (https://openai.com/business-data/). Recorded as Australian region available.
Ask: Is our project set to Australia storage with MAM or ZDR, and do we accept out-of-region inference?
Endpoints
Public claim only
Chat/completions and responses are ZDR-eligible with limits. Assistants, threads, vector stores, files, fine-tuning, batches, evals, agents, videos are not ZDR-eligible and keep application state until deleted (assistants objects 30 days after API deletion, otherwise indefinitely). Web search and remote MCP tools are listed for AU storage.
Ask: Which endpoints and tools are enabled, especially files, vector stores and web search?
Australia
Public claim only
AU storage is documented. AU processing is not. Astra for Law's corpus is US case law/statutes. Trusted Access is described for eligible law firms (US-centric announcement).
Ask: Is Astra for Law licensed for an Australian API org, and does the index include any Australian authorities?
Data, controls and requirements
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Client data
Not confirmed
Trusted Access: eligible lawyers and people working under their supervision; ZDR on the API for eligible firms; ChatGPT Enterprise usage excluded from human review by default (that last point is ChatGPT, not the API default). The wrapping app may still store prompts.
Ask: Is our use under Trusted Access, and what does the wrapping app retain?
OpenAI models
Public claim only
OpenAI is the model provider. Astra for Law = GPT-6 Astra plus legal search index and instructions; API id gpt-6-astra-law (coming soon). Hosting subprocessors are not substitute model vendors.
Ask: Pin model snapshots in production and whether any request is routed elsewhere.
Subprocessors
Public claim only
Published list at openai.com/policies/sub-processor-list/. Regional processing uses Cloudflare Regional Services on us/eu prefixed domains. Microsoft, AWS, GCP, Oracle, CoreWeave, Cerebras appear as infrastructure.
Ask: Which subprocessors apply to our region and endpoints, with notice rights?
No training unless opt-in
Public claim only
Services Agreement 4.2: OpenAI will not use Customer Content to develop or improve the Services unless Customer explicitly agrees. Endpoint table: Data used for training = No across listed endpoints. API dashboard opt-in (feedback/evals) is the exception. Fine-tuned models are not shared with other customers (Enterprise sibling; confirm for our org).
Ask: Confirm the org has no data-share opt-in and that fine-tunes stay private.
Playground opt-in
Public claim only
API customers can opt in by providing feedback in the Playground or evals. Default is opted out.
Ask: Is playground/eval sharing off for production keys?
30-day logs vs ZDR
Public claim only
Default abuse-monitoring logs up to 30 days. ZDR or MAM (prior approval) exclude customer content from those logs, with image/file exceptions. ZDR forces store=false. Responses API otherwise stores ~30 days when store=true. Assistants/files/vector stores retain until deleted (assistants 30 days after delete, else indefinitely). Videos blocked for MAM/ZDR.
Ask: Is the production project ZDR, MAM or default 30-day, and which stateful endpoints are used?
Deletion
Public claim only
Customer must delete application-state objects via API/dashboard. Assistants-related objects: 30 days after delete. DPA deletion/return applies to the signed copy. No public deletion certificate.
Ask: Who deletes files/vector stores/fine-tunes on matter close, and within what SLA?
Human review
Public claim only
Default abuse monitoring may involve review of flagged content. ZDR/MAM: OpenAI may still retain and human-review content on models made ineligible for severe-risk investigation, with written notice. Astra Trusted Access: ChatGPT Enterprise usage excluded from human review by default — that is not the unmodified API default.
Ask: Is Eyes Off / Safety Retention / Trusted Access provisioned for our org ID?
Org and project
Public claim only
Organisation- and project-level retention controls. Admin API and Audit Logs API advertised. Key hygiene is the customer's job.
Ask: Show project isolation, key rotation and that production is ZDR.
Encryption
Public claim only
Business-data / Edu pages state AES-256 at rest and TLS 1.2+. Confirm for API in the DPA/trust reports.
Ask: Encryption and CMK for regional storage projects?
Assurance
Public claim only
CSA STAR, SOC 2 Type 2, ISO 27001/27017/27018/27701 on the business-data page covering the API platform. Reports not inspected.
Ask: Current reports scoped to the API, including ZDR.
Services Agreement
Public claim only
Services Agreement 4.2 no-training unless explicit agreement; 4.3 customer evaluates Output. DPA at openai.com/policies/data-processing-addendum/. ZDR/MAM require additional approved terms. Astra Trusted Access is a separate eligibility programme.
Ask: Signed SA, DPA, ZDR addendum and Trusted Access terms, with liability and law/forum.
AU store / process elsewhere
Public claim only
Australia: storage yes, processing no. Content may be processed and temporarily stored outside AU to deliver the service. us/eu prefixed domains keep TLS termination in-region via Cloudflare.
Ask: APP 8 assessment of AU storage with non-AU inference.
Incidents
Public claim only
DPA exists (sibling Enterprise scrape: without undue delay). Confirm the executed API DPA clock.
Ask: Numbered notice period and NDB cooperation?
Changes
Public claim only
OpenAI may make models ineligible for ZDR/MAM with prior written notice. Regional support tables change. New endpoints may store state even under org-level ZDR.
Ask: How are we notified of ZDR-eligibility or regional-table changes?
Exit
Not confirmed
Customer-controlled deletion of files, vector stores, fine-tunes and assistants. Application copies in our systems are separate. DPA return/deletion on the signed copy.
Ask: Runbook to revoke keys, delete stateful objects and certify provider deletion.
Tools
Public claim only
Responses File Search, Web Search, Code Interpreter, Remote MCP server tool are listed even for AU storage. Third-party tools follow their own terms and may retain state. ZDR does not make those endpoints ZDR-eligible.
Ask: Which tools are enabled, and is web search/MCP off for privileged matters?
Web search tool
Public claim only
AU regional table lists /v1/responses Web Search as a storage-supported service. Astra for Law uses a legal search index plus (in the bench) comparison against web search. Tool output leaves the closed prompt.
Ask: Is the web-search tool disabled on production projects?
Agents and MCP
Not confirmed
/v1/agents is not ZDR-eligible and stores state until deleted. Remote MCP is listed. The wrapping application decides whether tools can write to DMS or email.
Ask: Which agent/MCP tools can take external actions, and with whose approval?
Oversight
Public claim only
SA 4.3: customer is solely responsible for evaluating Output. Astra help (ChatGPT) still requires review of answers and cited sources. No API-enforced lawyer gate.
Ask: What evaluation and logging does our application enforce before a lawyer relies on output?
