Security, privacy, and data controls

The official, up-to-date statement lives in the Enterprise privacy page and the API data usage docs. Read them once; they change rarely but they are the source of truth.

"Not used for training" is not the same as "not stored"

Even when your data is excluded from training, it still passes through and is usually retained for a limited window for abuse monitoring and operations. On the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète →, standard retention is around 30 days, then deletion, unless you qualify for Zero Data Retention (ZDR), which OpenAI grants to eligible enterprise use cases so that nothing is stored after the request completes.

The takeaway: "won't train on it" answers the IP question. "How long is it stored and who can see it" is a separate question you answer with retention settings and ZDR.

Where data actually goes when you use Actions and Connectors

This is where people get surprised. ChatGPT is not a sealed box once you add tools.

Connectors

A Connector (Google Drive, SharePoint, GitHub, Box, and others) gives ChatGPT scoped, read-style access to an external system so it can pull context into a chat or power deep research. Two things to internalize:

1. The data flows both ways at query time. When ChatGPT searches your Drive to answer a question, snippets of those files enter the conversation. They are now governed by *your account's* data policy, not the source system's.

2. Admins control which Connectors exist. In Enterprise and Team workspaces, a workspace admin enables Connectors centrally. If you are setting this up for a team, restrict Connectors to the systems you actually vetted.

See the current list and behavior in the ChatGPT Connectors help docs.

GPT Actions

An Action lets a Custom GPT call an external APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → you define via an OpenAPI schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.Voir la définition complète →. This is the riskiest surface because *you* are wiring the data path. When a user talks to your GPT, ChatGPT can send conversation content to whatever endpoint your Action points at.

Three controls matter:

• Auth: Use OAuth or APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète →-key auth in the Action config, never hardcode a secret into instructions.
• Scope: The endpoint should only do what the GPT needs. A "look up order status" Action should not expose a "delete customer" route.
• Data minimization: Your Action receives whatever the model decides to send. Validate and sanitize on your server.

Here is a minimal, safe Action schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.Voir la définition complète → fragment. Note the explicit auth and the read-only single purpose:

The pattern constraint is a small but real control: it stops the model from sending malformed or injected IDs to your backend.

The ChatGPT agent, Codex, and scheduled tasks

The ChatGPT agent can browse, click, and run actions on your behalf, which means it can submit forms and move data autonomously. Codex operates over code repositories. Scheduled tasks run prompts on a timer, sometimes while you are away.

The risk profile rises with autonomy. An agent that can act on a connected system can be steered by prompt injection: malicious text hidden in a web page or document that tries to hijack the agent's instructions. Treat any autonomous run that touches sensitive systems as something that needs a human approval step, and prefer read-only access for anything experimental.

A note on memory, custom instructions, and Projects

These three features quietly persist data across chats, so they deserve a security pass.

• Memory stores facts the model learns about you across conversations. If you paste a client's data into a chat, memory could retain a summary of it. Review and clear memory in Settings → Personalization → Memory. In Enterprise workspaces, admins can govern this.
• Custom instructions are sent with every message. Do not put secrets or full client names in them.
• Projects keep files and chats scoped to one workspace. This is actually a *good* control: it contains a client's documents to one Project instead of leaking them across your whole history. Use Projects as containment boundaries.

The custom GPT and GPT store exposure

If you publish a Custom GPT to the GPT Store, remember what is shared and what is not.

• Your instructions and configuration can be partially inferred by determined users; do not treat the system prompt as a secret vault.
• Uploaded knowledge files attached to a GPT can sometimes be extracted by users through clever promptingpromptingPrompt engineering is the practice of designing and refining text inputs to guide large language models toward accurate, relevant, and reliable outputs.Voir la définition complète →. Never attach a file to a public GPT that you would not be comfortable handing to a stranger.
• Conversations users have with your GPT are governed by *their* account settings and OpenAI's policies, not yours.

For internal-only GPTs, set sharing to "Only people in my workspace" rather than public. That one setting is the difference between a private tool and a public one.

A simple rule for safe use at work

You do not need a 40-page policy. You need one rule that holds up under pressure:

> If the data is regulated, secret, or someone else's, it only goes into a workspace that contractually excludes training and that an admin controls. Everything else can go in your personal account.

Unpack it:

• Regulated: health, financial, personal data covered by GDPR, HIPAAHIPAAHealth Insurance Portability and Accountability Act, loi américaine imposant la protection des données de santé (PHI). Violations : amendes jusqu'à 1,9M$ par catégorie de violation.-style rules. Business workspace only, and check your DPA.
• Secret: unreleased products, source code, strategy, credentials. Business workspace, and never in custom instructions or public GPTs.
• Someone else's: client data, partner data, anything you are contractually bound to protect. Business workspace, ideally scoped to a Project.

If none of those three apply (you are drafting a blog post, learning a concept, summarizing a public article), your personal account is fine.

Applying the rule to the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète →

On the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → the same principle holds, plus you own the system prompt and retention configuration. Use structured outputs to constrain what the model returns so a downstream system never receives free-form text it might mishandle. Here is the shape of a request that forces a strict schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.Voir la définition complète →:

from openai import OpenAI

client = OpenAI()

resp = client.responses.create(
    model="gpt-4.1",
    input="Extract the order ID and status from: 'Order AB12CD34 shipped today.'",
    text={
        "format": {
            "type": "json_schema",
            "name": "order",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {
                    "order_id": {"type": "string"},
                    "status": {"type": "string"},
                },
                "required": ["order_id", "status"],
                "additionalProperties": False,
            },
        }
    },
)

print(resp.output_text)

strict: True plus additionalProperties: False guarantees the model cannot smuggle extra fields into your pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète →. That is a privacy control as much as a correctness one: it bounds exactly what data crosses your system boundary.

If you build agents

When you use function calling or the Agents SDK to give a model tools, every tool is a data egress point. Apply least privilege to each function the same way you would to the YAML Action above: narrow scope, validated inputs, and a human checkpoint before any irreversible or sensitive action. The model deciding *when* to call a tool does not absolve your code of validating *what* it calls the tool with.

Quick admin checklist for a team rollout

If you are the person turning ChatGPT on for colleagues:

1. Use a Team or Enterprise workspace so training exclusion is contractual, not per-user.

2. Enable only vetted Connectors, and prefer read-only scopes.

3. Set Custom GPT sharing defaults to workspace-only.

4. Decide your memory and retention posture, and ask about ZDR if you handle regulated data.

5. Publish the one-sentence rule above to every user. Simplicity is what makes a policy actually followed.

Key Takeaways

• Tier determines training by default. Personal accounts may train on your data unless you opt out; Team, Enterprise, Edu, and the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → exclude it contractually. Confirm in the Enterprise privacy page.
• "Not trained on" is not "not stored." Ask about retention windows and Zero Data Retention separately when data is sensitive.
• Actions and Connectors are data exits. Give each one narrow scope, real auth, and validated inputs; treat autonomous agents and scheduled tasks as higher risk and add human approval.
• Use Projects and workspace-only sharing as containment. They keep a client's files from leaking across your history or to the public GPT Store.
• Carry one rule: regulated, secret, or someone else's data goes only into an admin-controlled workspace that excludes training. Everything else is fine in your personal account.