Can an AI agent delete its account? 59 products

Our agent deleted its own account on 40 of 59 products (1–29 Oct 2026), only 6 over the API. What blocked it and what to change.

By AgentDrives LabPublished 30 October 20265 min readTopic Cancel and leaveFor agents Markdown

Supabase, Clerk, GitHub

Yes. Our agent deleted the account it opened on its own on 40 of 59 products, driven 1 to 29 Oct 2026. On 6 of those 59 it did so over the API; on 34 in the account settings in a browser.

Key takeaways

  • On 19 of 59 it could not delete the account on its own.
  • On 7 of those 19 it could remove its keys and data, but the login stayed.

If agents sign up for your product, the way out is a step to design, the same as the way in. Our advice: put deletion where the agent already works, and make sure the old key stops working.

What our agent ran into at the leave step

Our agent reached the delete step on 59 of the 191 products it drove.

Outcome Products (of 59)
Deleted over the API 6
Deleted in the account settings in a browser 34
Could not delete on its own 19
Total that reached the delete step 59

Of the 40 products where the account was deleted, 37 refused the old key or login afterwards. Of the 19 where it could not be deleted, 7 let our agent remove its keys and data but kept the login.

Deleted is not the same as gone

A complete exit has three parts:

  • The login is refused. After deletion the user can no longer log in (zitadel.com[1]).
  • The data goes with the account. Apple asks apps to “Offer to delete the entire account record, along with associated personal data.” (developer.apple.com[2]) Google Play says the user data tied to the account must be deleted along with it (support.google.com[3]).
  • Every token stops working. OAuth 2.0 Token Revocation (RFC 7009) adds an endpoint to revoke tokens (rfc-editor.org[4]), and an IETF draft on AI agent authentication asks that revoked authorization be enforced without undue delay (datatracker.ietf.org[5]).

Removing keys and data but keeping the login is a partial exit

Apple asks apps to offer deletion of the entire account record, not only deactivation (developer.apple.com[2]). On 7 of the 19 products where our agent could not delete its account, it could clear keys and data but the login stayed.

Exit level What ends What stays
Token revoked The token (RFC 7009 revokes tokens, not accounts) (rfc-editor.org[4]) The account and its login
Keys and data removed Keys and data The login
Account deleted The account record and its personal data (developer.apple.com[2]) Nothing that lets the old login in (zitadel.com[1])

What the rules require, and where they stop

None of these texts speaks of accounts an agent opened; they speak of people, apps and personal data.

Rule What it requires
Apple App Store Apps that support account creation must let people start deletion within the app (developer.apple.com[2]). If deletion finishes on a website, the app links directly to that page (developer.apple.com[2]). Outside highly regulated industries, no phone call, email, chat or other support flow (developer.apple.com[2]). A manual process is fine if the person is told how long it takes and gets a confirmation (developer.apple.com[2]).
Google Play Apps that allow in-app account creation must let users request deletion (support.google.com[3]). The user data goes with the account (support.google.com[3]).
GDPR Art. 17(1) Erasure of personal data without undue delay, where a listed ground applies (gdpr-info.eu[6]).
GDPR Art. 12(3) Report action taken on a request within one month of receipt (gdpr-info.eu[7]).

Should an API key be able to delete its own account?

It can, if the delete call is labelled as destructive and a confirm step sits in front of it. That is our advice; the specs give you the pieces.

MCP tool annotations include a destructiveHint flag: “If true, the tool may perform destructive updates to its environment.” (raw.githubusercontent.com[8]) Those annotations are hints, not guarantees of what a tool does (raw.githubusercontent.com[8]). The MCP tools spec says a human should be in the loop, able to deny tool invocations, and that clients should “Prompt for user confirmation on sensitive operations” (modelcontextprotocol.io[9]).

Identity providers draw the line in different places. Supabase’s deleteUser needs a service_role key and should run on a server, not with the user’s own session (supabase.com[10]). Clerk lets users delete their own accounts from its profile component, once the app turns the feature on (clerk.com[11]). GitHub releases a deleted account’s username to others after 90 days (docs.github.com[12]).

A decision guide, by how agents reach your product:

  • Your agents use an API key. Offer a delete endpoint that is irreversible and says so, as Auth0’s does (auth0.com[13]). Our advice: require an explicit confirmation in the call.
  • Your agents use your MCP server. Offer a delete tool flagged with destructiveHint, so the client can ask its human before it runs (raw.githubusercontent.com[8], modelcontextprotocol.io[9]).
  • Your agents use the browser. Give a direct link to the page where deletion completes, as Apple asks of apps (developer.apple.com[2]).

What to change in your product

Put the delete step where the agent already works, and check that the old key and login are refused once it is done. Our checklist:

  1. A delete endpoint, or an MCP tool flagged with destructiveHint, with a confirm step (raw.githubusercontent.com[8], modelcontextprotocol.io[9]).
  2. A direct link to the page where deletion completes (developer.apple.com[2]).
  3. No path that runs through a support ticket, email or chat (developer.apple.com[2]).
  4. Say how long deletion takes and confirm when it is complete (developer.apple.com[2]).
  5. Revoke every token and refuse the old login, without undue delay (rfc-editor.org[4], datatracker.ietf.org[5]).
  6. Delete the login along with the keys and data, not only the keys and data (developer.apple.com[2]).

Step by step: how to let agents delete the account they opened. For the wider exit, including cancelling a plan, see whether an agent can cancel and leave on its own, and for the ways in that agents use, which agent surfaces products offer, and how many worked.

How we measured

From 1 to 29 Oct 2026 our agent drove 191 products and got far enough to try deleting its account on 59. On each we recorded whether our agent deleted its account on its own over the API, through the account settings in a browser, or not at all, and whether the account was really gone afterwards, meaning its old key or login was refused. The scope is these 59 products in that window. Full method: how we drive products.

See what your docs tell an agent about leaving.

Free: what an agent finds, any wall in its way and how many things to fix. The Audit ($499) gives you every fix.

Pages this guide cites.

As linked when it was written. ↩ goes back to where the text cites it.

  1. 1zitadel.comdocs/apis/resources/user_service_v2/user-service-delete-user↩
  2. 2developer.apple.comsupport/offering-account-deletion-in-your-app↩
  3. 3support.google.comgoogleplay/android-developer/answer/13327111↩
  4. 4rfc-editor.orgrfc/rfc7009↩
  5. 5datatracker.ietf.orgdoc/draft-klrc-aiagent-auth/02↩
  6. 6gdpr-info.euart-17-gdpr↩
  7. 7gdpr-info.euart-12-gdpr↩
  8. 8github.com · raw.githubusercontent.commodelcontextprotocol/modelcontextprotocol/blob/main/schema/2025-06-18/schema.ts↩
  9. 9modelcontextprotocol.iospecification/2025-06-18/server/tools↩
  10. 10supabase.comdocs/reference/javascript/auth-admin-deleteuser↩
  11. 11clerk.comchangelog/2023-06-23↩
  12. 12docs.github.comen/account-and-profile/reference/personal-account-reference↩
  13. 13auth0.comdocs/api/management/v2/users/delete-users-by-id↩

Asked about this guide.

7 questions
Can an AI agent delete the account it signed up for without a human?

Yes. With no person involved, our agent removed the account it had opened on 40 of the 59 products where it reached the delete step, in drives from 1 to 29 Oct 2026.

Should account deletion be exposed over the API, or is a settings page enough for agents?

A settings page worked for our agent on 34 of 59 products, against 6 of 59 over the API. Our advice is to offer both: an API path for agents that never open a browser, and a direct link to the deletion page, the way Apple asks apps to link straight to it.

Is it a risk to let an API key delete the account it belongs to?

Auth0 notes its own delete call cannot be undone. MCP lets a tool mark itself destructive and says clients should ask for confirmation on sensitive operations. Those flags are hints only. Supabase keeps its delete behind an admin service role key on the server. Our advice: pair the key's delete right with a confirm step.

After an agent deletes its account, are its API keys and login actually refused?

Not on every product. The old key or login was refused on 37 of the 40 products where our agent deleted its account. The end state to aim for is the one ZITADEL documents: the user can no longer log in.

Is deleting API keys and data enough, or does the login also need to go?

The login needs to go too. Apple counts only deactivating or disabling an account as insufficient and asks for the whole account record to be deleted. Google Play ties the user data to the account deletion. In our drives, 7 of the 19 products that blocked a full deletion let the agent clear keys and data while the login stayed.

As an agent builder, how do I know before signing up whether my agent will be able to leave a tool?

Look for a delete endpoint in the API docs, or an MCP tool marked as destructive. The [short answer on agents deleting the account they opened](https://agentdrives.com/access/leave/delete-account) is kept up to date.

Do privacy rules or app store rules require that accounts can be deleted, and does that cover agent-created accounts?

Apple and Google Play require apps with account creation to let users start or request deletion. Outside highly regulated industries, Apple rules out deletion that runs only through a support flow. GDPR gives a right to erasure of personal data, a one-month limit to report action, and asks for electronic means to make requests. None of these texts addresses accounts an agent opened.

Cite this

AgentDrives Lab. “Can an AI agent delete its account? 59 products.” 30 October 2026. agentdrives.com/guides/can-an-ai-agent-delete-its-own-account

Drive data is CC BY 4.0. The text as Markdown. Every fact with its evidence.

More from the drives.