Is it safe to put client code into GitHub Copilot?
This app has not yet been installed and tested. Treat the record below as provisional.
GitHub Copilot
Free / Pro / Pro+ / Max, individual · tested 10 Aug 2026- Input retention
- Not stated as a single period. GitHub says it maintains zero data retention agreements with OpenAI, Anthropic and xAI for the models it serves. One documented exception: for the Claude Fable 5 model, Anthropic retains prompts and outputs to operate safety classifiers. Docs read 10 Aug 2026.
- Used for training
- Yes, by default, on individual plans. GitHub may use 'interactions with GitHub features and services—including inputs, outputs, code snippets, and associated context' to train models. Effective 24 Apr 2026. Business and Enterprise are excluded.
- Opt-out
- Yes. Setting is 'Allow GitHub to use my data for AI model training', set to Disabled. Individual plans only — it is enabled by default, so this is an opt-out and not an opt-in.
- Sub-processors
- Model providers named in GitHub's own docs: OpenAI, Anthropic, Google, xAI. Hosting spans GitHub's Azure infrastructure, OpenAI, Anthropic PBC, Amazon Web Services and Google Cloud Platform.
- Consumer vs business
- The whole story. Individual plans train on your data unless you turn it off. Copilot Business and Copilot Enterprise are excluded and covered by GitHub's Data Protection Agreement.
- Provider
- GitHub — contracting entity not confirmed against the agreement in this pass
- Jurisdiction
- Unverified
- Data residency
- Unverified — not established in this pass
- Known incidents
- None found in primary sources during this pass, 10 Aug 2026
On Copilot Free, Pro, Pro+ and Max, 'Allow GitHub to use my data for AI model training' is enabled by default as of 24 April 2026. Business and Enterprise are excluded. Check that toggle before you open client code.
The short answer
If you are on an individual Copilot plan — Free, Pro, Pro+ or Max — GitHub’s own documentation says your inputs, outputs and code snippets may be used to train models, and the setting that controls it is enabled by default. That has been the stated position since 24 April 2026.
If you are on Copilot Business or Copilot Enterprise, you are excluded from that and covered by GitHub’s Data Protection Agreement.
For work under an NDA or a client contract that restricts third-party processing, the individual plans are not a safe default and the enterprise tiers are the ones to be on. This is a plan question far more than it is a vendor question.
What the documentation actually says
The setting is named, in GitHub’s own words, “Allow GitHub to use my data for AI model training”, and the documented instruction is to select it and click Disabled. It applies to Copilot Free, Pro, Pro+ and Max, with the policy dated 24 April 2026.
What it covers is broader than the phrase “your code” suggests. GitHub describes the scope as “interactions with GitHub features and services—including inputs, outputs, code snippets, and associated context”. The last two words are the important ones, and they are the reason what your IDE extension sends that you didn’t type matters here: the context an assistant assembles is not limited to the file you are looking at.
Business and Enterprise customers are excluded from this arrangement, with their data protected under GitHub’s Data Protection Agreement rather than by a toggle.
Which companies actually see your code
Copilot is not one model. GitHub’s model hosting documentation names four providers and describes a different data posture for each — which means “is Copilot safe” has four answers depending on which model you selected in the picker.
| Provider | Where GitHub says it is hosted | Stated retention position |
|---|---|---|
| OpenAI | OpenAI and GitHub’s Azure infrastructure | Zero data retention agreement |
| Anthropic | AWS, Anthropic PBC, Google Cloud Platform | Zero data retention agreement for generally available features — with one exception below |
| Google Cloud Platform | Gemini “doesn’t use your prompts, or its responses, as data to train its models” | |
| xAI | xAI infrastructure | Zero data retention API policy; content exists “only temporarily in RAM” |
The exception is the most useful finding on this page. GitHub’s documentation records that for the Claude Fable 5 model, Anthropic retains prompts and outputs in order to operate safety classifiers — a carve-out from the zero retention agreement that applies to Anthropic’s generally available features.
That is not hidden, and it is not unreasonable. It is simply the kind of detail that a blanket “we have zero data retention agreements with our model providers” summary does not carry, and it changes the answer for anyone whose contract prohibits third-party retention outright.
Disclosure: this site is drafted with Claude, made by Anthropic, and this page assesses Anthropic as one of Copilot’s model providers. That is a conflict of interest. Verify the retention position independently against GitHub’s and Anthropic’s own documentation — including against this page. See the Anthropic dossier and how this site makes money.
Where the default bites
Three situations turn this from a preference into a problem.
Consultants and agencies. If your client contract restricts processing their code by third parties, the individual-plan default puts you in breach from the first completion, and the plan you are on is the fact that determines it. Can you use AI coding tools under an NDA? covers the analysis.
Organisations paying by expense claim. An engineer expensing a personal Pro subscription is operating under individual terms, with no DPA, on a default that trains. From a compliance perspective this is invisible until someone looks — which is exactly the pattern in shadow AI: finding unapproved tools.
Anyone who enabled Copilot before April 2026. The policy has a date. A setting you reviewed in 2025 is not evidence about the state of your account today, and defaults that change are the reason to re-check rather than remember.
What to change today
- Confirm which plan you are actually on. Not which plan your employer buys — which one your account uses. Individual and organisation seats look similar in the editor and are governed by different terms.
- On any individual plan, set “Allow GitHub to use my data for AI model training” to Disabled. This is the single action on this page.
- Check which model is selected. The retention position differs per provider, and the Fable 5 carve-out above applies to one specific model.
- Add ignore rules before you open a client repository, so the “associated context” GitHub describes does not include credentials or files you never intended to send. Our IDE ignore file generator produces a starting configuration, and how to stop secrets leaking into an AI assistant explains why the ignore file needs testing rather than trusting.
- For any regulated or contracted work, move to Business or Enterprise. The exclusion there is contractual rather than a toggle a user can flip back.
If you are under GDPR, ISO 27001 or PCI DSS
The DPA is the dividing line, and it is available on Business and Enterprise rather than on individual plans. Without one you have no Article 28 relationship for anything personal that reaches the tool, which is a documentation problem before it is a security problem — see how to get a DPA for an AI tool.
Two fields in the record above are honestly Unverified: the contracting legal entity and data residency. Both are answerable from the customer agreement and from GitHub’s enterprise documentation, and neither was established in this pass. They are not marked “none” or guessed at, because an unverified field is a publishable value and an invented one is not.
What this record is, and is not
This is desk research read on 10 August 2026 against GitHub’s own published documentation. Nobody has sat with a network proxy watching what the extension transmits, and the page carries a “desk research only” badge for that reason.
The claims here are GitHub’s claims. What a vendor commits to in documentation and what an extension does on the wire are different questions, and the second one needs a hands-on test that has not happened yet.
Corrections are welcome and get made quickly — tell me if a figure has moved.