ChatGPTOpenAIApple MessagesAI AgentsAI Governance

ChatGPT Can Send Your Messages. Treat Approval as an Identity Control

OpenAI's Apple Messages plugin moves ChatGPT from drafting into identity-bearing action. Operators need fresh approval, visible recipients, and revocable permissions.

IndieStudio •

OpenAI has added an Apple Messages plugin to Codex and ChatGPT Work on the macOS desktop app. On a compatible Apple Silicon Mac, the agent can read and search iMessage, SMS, and RCS conversations, summarise a thread, draft a response, and send it through Apple’s Messages app.

The feature is easy to frame as a productivity upgrade. Less searching, less copying, and fewer repetitive replies. But sending a message is not another text-generation task. It is an external action attributed to a person.

That distinction changes the control design.

Drafting is reversible. Sending is authority.

A draft can be inspected, edited, or discarded. A sent message reaches a customer, employee, partner, friend, or family member with the apparent authority of the account holder.

If the agent misunderstands the context, selects the wrong recipient, or follows a misleading instruction inside the conversation, the mistake leaves the private workspace immediately.

OpenAI’s default is sensible. According to its Apple Messages plugin documentation, the user approves the message and its recipients before a send. An Allow once option grants permission for that single action. The user can also choose persistent approval for a conversation, which lets ChatGPT send future messages there without another send confirmation.

OpenAI’s own warning is the important part: persistent approval removes the final chance to review a message before ChatGPT sends it as the user. The documentation recommends keeping per-send approval for conversations that may contain untrusted or misleading instructions.

This is a clean example of the difference between capability and authority. The model may be capable of producing a reasonable reply. That does not mean it should have standing authority to publish the reply through a trusted identity.

Imagine giving an assistant access to company letterhead and a signature stamp. Asking it to prepare a response is ordinary delegation. Letting it choose the recipient, final wording, and moment of delivery without another check is a different level of responsibility.

Classify conversations by consequence

Customer commitments, employee matters, legal discussions, payments, health information, and account recovery should keep fresh approval for every send.

Low-risk household logistics or a tightly scoped internal test thread may justify a limited persistent grant, but only when the user understands the failure mode. Convenience is not a risk classification.

The same principle applies to group conversations. A familiar thread can change when someone adds a participant. Persistent permission should not silently survive a material change in the audience.

Make approval meaningful

Show the final message and every recipient together. A confirmation screen that hides group membership or truncates the wording creates approval theatre. The human needs enough context to catch a plausible message aimed at the wrong audience.

The approval should also be fresh. Reusing yesterday’s permission for today’s different context turns a deliberate decision into background access.

For consequential conversations, require the reviewer to see:

  • the complete recipient list
  • the final, untruncated wording
  • any attachment or link that will be sent
  • the action the agent will take next

Keep standing permission visible and temporary

Persistent grants should be easy to list, review, and revoke. Teams should know which conversations an agent can currently write into, who granted that authority, and when it was last used.

A permission that cannot be inspected becomes forgotten authority. Add expiry dates to standing grants where possible, and review them after role changes, device changes, or incidents.

Revocation also needs testing. Removing access in a settings screen is not enough if the agent can still act through a cached session or another connected device.

Log the action path

Record what the agent proposed, which recipient set was shown, who approved the send, and what was delivered. The point is not surveillance. It is accountability and incident reconstruction when something goes wrong.

There are still unanswered privacy questions. TechCrunch’s launch coverage raised questions about how message data is handled. OpenAI says the plugin runs locally and does not create an index of every message, but the reviewed material does not provide a complete data-flow diagram or retention contract. Operators should validate workspace data controls before connecting sensitive message history.

The operator takeaway

Messaging agents are useful when they remove typing, searching, and repetitive preparation. They become risky when that convenience quietly turns into standing authority.

Keep per-send approval for consequential conversations. Show recipients and wording together. Limit and expire persistent grants. Log approved actions. Test revocation before relying on it.

Let the agent prepare the message. Keep a human responsible for the moment that message becomes yours.