GPT WorkspaceTrust Center
Where your data goes, and where it does not.
Every processor that touches a request, what each one receives, how long anything is kept, and the controls an enterprise can turn on. Written for security reviews and procurement, kept in sync with the code.
ISO 27001
Certified, audit report public
GDPR
Art. 28 DPA on request
CASA Tier III
App Defense Alliance
No file storage
Your files stay on Google's servers
Data flow
One request, step by step
You type a prompt in the chat, the sidebar, a Sheets formula or an MCP client. GPT Workspace reads the Google file you point at with your own Google token, builds the request on Google Cloud, and sends it to the model through OpenRouter. The answer comes back the same way and is written to your Google file or shown in the chat. Nothing from the file is kept once the request ends.
1. Your Google account
Files, e-mails and sheets are read in real time with your OAuth token, scoped to what you allowed.
2. GPT Workspace on Google Cloud
Builds the prompt, checks your plan, logs usage metadata (counts, dates, document names). Does not copy the file.
3. OpenRouter
The single gateway to every model. Carries the request to the provider pinned for that model.
4. The model provider
Generates the answer under API terms that forbid training. The provider's retention applies to the request.
Text model providers
What each provider receives is the prompt, the context you attached (file excerpts, the selected range, the e-mail thread) and the tool calls of the request. Which model answers depends on the feature and your plan; the list below is every provider that can be involved.
| Provider | Models | Used for | Endpoint routing | Training on your data |
|---|---|---|---|---|
| OpenAI | gpt-5.4, gpt-6-luna | Chat, agents, Sheets formulas, slide planning | gpt-6-luna is pinned to OpenAI's own endpoint with fallbacks off; gpt-5.4 uses OpenRouter's default routing for that model | API terms forbid training on the requests (openai.com/policies/api-data-usage-policies) |
| Anthropic | Claude Sonnet, Claude Opus (latest) | Slide design and rendering, advanced chat | Pinned to Anthropic's own endpoint, fallbacks off: never served by a third-party host | Commercial terms forbid training on the requests (anthropic.com/legal/commercial-terms) |
| Google (AI Studio) | Gemini 3.5 Flash Lite, Gemini Flash (latest) | Chat, style guides, fast agent steps | Pinned to Google AI Studio, fallbacks off | Paid API terms forbid training on the requests (ai.google.dev/gemini-api/terms) |
| Perplexity | Sonar, Sonar Pro Search | Web search in chat and the GPT_WEB formula: the search query and the conversation context needed to answer it | OpenRouter's default routing for Perplexity models | API terms forbid training on the requests (perplexity.ai/hub/legal/perplexity-api-terms-of-service) |
| Z.ai (GLM-5.3-flash), served by OpenRouter inference hosts | GLM-5.3-flash | The default agent orchestrator, Sheets formulas (standard and advanced tiers) | Open-weight model served by pinned OpenRouter inference hosts | Inference hosts only: the model author never receives the prompts |
| OpenAI gpt-oss-120b, served by OpenRouter inference hosts | gpt-oss-120b | Lightweight tasks | Open-weight model, OpenRouter's default routing | Inference hosts only |
| Typesafe | jev-1.13 (decision model) | The GPT_CLASSIFY formula: the text to classify and the candidate labels | OpenRouter, pinned to Typesafe | API terms forbid training on the requests |
Image models
| Provider | Models | Used for |
|---|---|---|
| OpenAI (via OpenRouter) | GPT Image 1 mini, GPT Image 2 | Image generation in chat, slides and the GPT_IMAGE formula: the prompt and, for edits, the source image |
| Google (via OpenRouter) | Gemini 3.1 Flash Lite Image, Gemini 3 Pro Image | Default image model for the same features |
Web search
When you ask the chat to search the web, or use the GPT_WEB formula, the search query and the conversation context needed to answer it go to Perplexity (Sonar) through OpenRouter. Stock images for generated slides are found with Google Programmable Search; only the image query is sent.
File search (attached files and custom assistants)
A file you attach to a chat, or add to a custom assistant, is uploaded to an OpenAI vector store so the model can search inside it. This is the one place a copy of a file leaves Google: the search index is dropped 1 day after it was last used and the uploaded copy is deleted after 7 days. Do not attach files that may not leave Google if that matters to you, and use the Drive agent, which reads in place, instead.
Retention
What we keep, where, and for how long
The figures below are the ones in our privacy policy. Deletion requests are handled within 24 hours.
| Data | Where | How long |
|---|---|---|
| Your Google files, documents, e-mails and spreadsheets | They stay on Google's servers. GPT Workspace reads them in real time to answer a request and does not copy them | Never stored by GPT Workspace |
| Account record: first name, last name, profile picture, e-mail | Firebase (Google Cloud) | 18 months after your last activity, or within 24 hours of a deletion request |
| Google access and refresh tokens | Firebase, encrypted at rest | Until you revoke access or delete the account |
| Chat threads, settings, usage metadata (number of prompts, dates, document names) | Firebase | 18 months after your last activity, or within 24 hours of a deletion request |
| Files attached to a chat or to a custom assistant (vectorized for file search) | OpenAI vector store | The search index is dropped 1 day after the thread or assistant last used it; the uploaded copy at OpenAI is deleted after 7 days |
| Prompts and answers sent to a model provider | The provider named in the tables above | Processed in real time; no long-term storage by GPT Workspace. The provider's own retention applies |
| Support e-mails | Google Workspace (Gmail) | Kept to process the request and improve the service |
| Invoices | Stripe | The duration required by tax law |
Subprocessors
Everyone who processes data on our behalf
The model providers in the tables above, plus the infrastructure and business services below. This list is Annex 3 of the DPA; we notify DPA customers before adding a subprocessor.
| Subprocessor | Purpose | Data | Location |
|---|---|---|---|
| Google Cloud Platform / Firebase | Hosting, database, file storage, secrets, authentication | Account record (name, e-mail, profile picture), encrypted Google tokens, chat threads, usage metadata, settings | United States (US East and US West regions) |
| OpenRouter | Routes every model request to the provider above | Prompts, attached context, tool calls and model answers, in transit | United States. Requests are logged for 30 days, only to handle support tickets, and are never used for training |
| OpenAI (file search) | Vector store for files attached to a chat and to custom assistants | The attached files themselves, uploaded to OpenAI | United States |
| Google Programmable Search | Stock image lookup for generated slides | The image search query only | |
| Stripe | Payments, invoices, billing portal | Billing name, e-mail, address, VAT number, payment method (held by Stripe) | United States and Ireland (Stripe, Inc. and Stripe Payments Europe), under the Stripe Data Processing Agreement |
| PostHog (self-hosted at analytics.qualtir.com) | Product analytics | Usage events, feature flags, the e-mail as the person id | European Union |
| Sentry | Error monitoring | Stack traces, request metadata, user id | European Union (EU data region) |
| Google Analytics | Website traffic (gpt.space only, not the app) | Page views, anonymized visitor data |
Security practices
How the platform is protected
ISO 27001
Information security management system certified by an independent auditor. The audit report is public: open it from the card above.
CASA Tier III
Google's Cloud Application Security Assessment (App Defense Alliance) at the highest tier, required for the Google Workspace scopes the app uses.
Google User Data Policy, Limited Use
Data obtained through Google Workspace APIs is used only to provide the features you see, never for advertising, never sold, and never read by humans except with your consent, for security, or as required by law.
Encryption
TLS 1.2+ for every connection. AES-256 server-side encryption at rest on Firebase, with key management through GCP IAM. Google access and refresh tokens are encrypted at rest.
Hosting and isolation
Google Cloud Platform, US East and US West. Every component that processes customer data runs inside the project's private network, behind load balancers and a web application firewall; per-user data is isolated by Firebase security rules.
Continuous monitoring
GCP Security Command Center scans every asset in scope for vulnerabilities and misconfigurations. Reports are available to enterprise customers on request.
Vulnerability disclosure
External security researchers are welcome to report vulnerabilities responsibly to contact@gpt.space; the vulnerability disclosure policy describes the process and our commitment to a timely, transparent response.
No training on your data
Model providers are used under API terms that forbid training on the requests. GPT Workspace never trains or fine-tunes a model on customer data.
Enterprise controls
Controls you can turn on today
The DPA and invoicing come with the Enterprise plan; deletion is available to every account.
Data processing agreement
A GDPR Article 28 DPA with the subprocessor list from this page, signed by GPTW LLC. Request it with the button below.
Deletion on request
Delete your account and every record we hold from the app, or by e-mail; handled within 24 hours.
Need a signed DPA or a security questionnaire?
Send your company name and the number of seats. We return the DPA pre-filled with the subprocessor list from this page, usually within two business days.
Data controller for the service: GPTW LLC, 150 Central Park, Los Alamos, New Mexico, USA.