On 13 August, OpenAI began rolling out a deeper Google Drive integration. Paid users on the web can pull a Doc, Sheet or Slide into a chat, keep it open beside the conversation, and edit it without leaving ChatGPT. The detail that matters is easy to skim past: ChatGPT edits the file stored in Drive rather than a separate copy. The rollout covers Plus, Pro, Enterprise, Edu, Healthcare and Business accounts, with mobile support promised later.

The mistake is filing this under “new feature”

Most teams will read the announcement, think “useful”, and carry on. The actual change is to the permission, not the interface. Reading and writing are different risks wearing the same OAuth grant.

When ChatGPT could only read, a bad output was a bad output. You saw it in the chat window, you judged it, you copied what you wanted. The human review step was structural. It happened because the tool physically could not put anything anywhere. Now the review step is optional, and optional review steps are the first thing a busy team drops in week three.

Most small teams do not have a content problem. They have a review problem. This change removes the last piece of friction that was quietly doing the reviewing for them.

Where this gets messy

Three practical failure points, in rough order of how likely they are to bite.

Shared drives are wider than people think. The Drive connection is not scoped to the one folder someone had in mind when they clicked through the consent screen. If a marketer connects a work account that has access to a shared client drive, the surface is the whole thing. For ChatGPT Enterprise and Edu there are admin controls, and a workspace admin can include or exclude specific shared drives. On individual Plus and Pro accounts, that governance layer does not exist. The person clicking approve is the entire approval process.

The undo is Drive version history, and nothing else. There is no diff view, no “here is what I changed” summary, no approval queue. If ChatGPT reworks a paragraph in a live campaign brief and three people read the new version before anyone notices, the correction is a manual restore from Drive’s revision list. That works. It is also the kind of thing nobody practises until the day they need it at 6pm.

Attribution disappears. An edit made by ChatGPT under a user’s account looks, in the version history, exactly like an edit made by that user. For an agency with client sign-off processes, that is a real problem. “Who approved this wording” is a question the audit trail can no longer answer honestly.

The documents you did not write

This is the part that gets almost no coverage, and it is the reason I would not connect a shared client drive today. Security researchers demonstrated at Black Hat in 2025 that instructions can be hidden inside a document, and that an assistant processing that file through a connector can act on them without the user doing anything. The technique was shown as a data leak: hidden text tells the model to find something sensitive and send it out.

Read access made that a disclosure risk. Write access makes it an integrity risk too. A marketing team pulls in a competitor PDF, a supplier price list, or a brief forwarded from a client’s client. If that file contains instructions the model treats as instructions, the model now has a hand on the real file.

I want to be careful here. This is a demonstrated technique, not a reported wave of attacks on marketing teams, and OpenAI has hardened connectors since. The honest framing is that the blast radius grew while most people’s habits stayed the same.

A simple version for small teams

You do not need a policy document. You need four decisions, made once, written somewhere findable.

• Decide which drive. Connect a personal or team drive that holds drafts, not the shared drive that holds signed contracts and client data. If your Drive access is all one bucket, that is the thing to fix first, and it is worth fixing regardless of ChatGPT.
• Keep client-facing files out of it. Draft in a working doc, move approved copy into the client-visible file by hand. Slower. Also the only version where sign-off means anything.
• Turn on the admin controls if you have them. On Enterprise and Edu the new Drive actions are off until an admin enables them, and shared drives can be included or excluded individually. Someone should actually make that choice rather than letting the default decide.
• Treat unfamiliar documents as untrusted input. If a file arrived from outside your organisation, read it yourself before handing it to an assistant that can write.

When this is genuinely worth it

None of the above means the feature is bad. For a solo founder maintaining their own content calendar in Sheets, or a two-person team iterating on landing page copy in a Doc nobody else touches, the time saved is real and the risk is close to zero. The copy-paste loop between a chat window and a document is genuinely tedious, and removing it is a good change.

The risk scales with how many other people depend on the file. That is the whole calculation. A doc only you read is a sandbox. A doc your client reads is production.

This is the kind of workflow mark8ng.ai is being built around: practical AI use with a human review step that survives contact with a deadline, rather than one that quietly gets skipped.

Before you connect anything this week, open your Google account permissions and look at what ChatGPT already has. Most people find at least one grant they forgot about, approved for a job that no longer resembles what the tool can do today.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post ChatGPT Can Now Edit Your Live Client Files. Nobody Re-Approved That. appeared first on Mark8ng.com.