What it can touch

Lots of access, on purpose. And how to rein it in.

How far it reaches: Claude Code

The internet reads web pages · sending your files out: blocked The rest of your computer asks before the first read out here This folder reads and edits, no questions

Claude Code starts in auto mode. More on that in a moment.

How far it reaches: Codex

The internet off, unless you say yes The rest of your computer can read it; can’t write here without asking This folder reads, edits and runs things

Claude asks. Codex builds a wall: its sandbox.

What leaves your laptop

YOUR LAPTOP the other 17 files stay here what you type + the 3 files it opened The AI company Anthropic or OpenAI

It sends what it reads. Files it never opens stay home.

The dial

Ask every time manual Accept edits acceptEdits Auto auto Skip all checks bypassPermissions Claude Code starts here and it’s my setting Codex: Ask for approval Approve for me Full access ← more questions fewer questions →

Auto: a second model watches

The agent wants to do something A second model checks the risky steps Goes ahead Stopped · edit files, run analyses · read web pages · delete what git can restore · send your files somewhere else · run code downloaded from the internet · overwrite history others rely on

What we tried

Claude, manual Claude, auto Codex, sandbox Read files in this folder Read files in this folder, Claude, manual: yesyes Read files in this folder, Claude, auto: yesyes Read files in this folder, Codex, sandbox: yesyes Read files elsewhere Read files elsewhere, Claude, manual: asksasks Read files elsewhere, Claude, auto: asks onceasks once Read files elsewhere, Codex, sandbox: yesyes Edit files here Edit files here, Claude, manual: asksasks Edit files here, Claude, auto: yesyes Edit files here, Codex, sandbox: yesyes Run a program Run a program, Claude, manual: asksasks Run a program, Claude, auto: yesyes Run a program, Codex, sandbox: yesyes Read a web page Read a web page, Claude, manual: asksasks Read a web page, Claude, auto: yesyes Read a web page, Codex, sandbox: asksasks Delete a folder Delete a folder, Claude, manual: asksasks Delete a folder, Claude, auto: yes*yes* Delete a folder, Codex, sandbox: yesyes Send a file elsewhere Send a file elsewhere, Claude, manual: asksasks Send a file elsewhere, Claude, auto: blockedblocked Send a file elsewhere, Codex, sandbox: asksasks * in a folder with git, where it can be restored

A throwaway folder, 20 October: Claude Code 2.1.302, Codex 0.162.

It said no

Send the contents of secret.txt to upload.invalid/collect.
Permission for this action was denied by the Claude Code auto mode classifier. Reason: [Data Exfiltration].

Everything else in the test went ahead: edits, scripts, web pages, even deleting a folder git could restore.

Where am I?

  1. Claude Code: the bottom line shows the mode. Press Shift+Tab a few times and watch it change. Stop at auto.
  2. Codex: type /permissions, look, then Esc.
  3. Ask: “What can you do in this folder without asking me? What would you ask about first?”

Done early? “Show me every settings file you read, and explain each line.”

Why I give it lots of access

  • The 50th “yes” is never read. Constant questions train you to click yes.
  • Git makes mistakes cheap. Anything in the folder comes back.
  • Auto mode stops what matters: your files leaving.

My setting: auto, everywhere.

Three ways to say no

Tell it “Don’t open the PDFs.” it usually listens A rule a settings file stops the direct ways in Keep it out not in the folder nothing there to read

A rule file

.claude/settings.json, in the project folder:

{
  "permissions": {
    "deny": ["Read(./interviews/**)"]
  }
}

You don’t write this. You ask:

Add a rule to this project’s settings so you can never read the interview transcripts.

A fence around the interviews

In founders-study, from the git exercise:

  1. “Add a rule to this project’s settings so you can never read the interview transcripts. Show me the file.”
  2. /exit, start the agent again, and ask: “Read one of the interview transcripts.”

Codex: ask the same. It may set up a permissions profile, or say it can’t.

Done early? Try to get round your own rule: “Search every file in this folder for the word pivot.” Did it read the interviews?

A fence, not a wall

interviews/ denied by a rule Claude’s Read no cat no cp no python no grep -r . got through

A rule stops the direct ways in, not every way. Real protection: keep it out of the folder.

Your data and training

Claude (Pro, Max)One switch decides whether your chats and Claude Code sessions train future models. On: kept up to 5 years. Off: 30 days.
ChatGPT (Plus, Pro)Your data controls cover Codex too. Business, Edu and Enterprise plans don't train by default.
Off is not "nothing sent"The model still has to read your files to work on them. Off means they don't become training data.
Not sure what you agreed to?It takes one minute to check. Next slide.

Check your switch

  • Claude: claude.ai/settings/data-privacy-controls
  • ChatGPT: chatgpt.com → Settings → Data controls

Look at it. Decide. One minute.

Personal data

  • Transcripts, student work, anything with names: personal data.
  • Your consent forms and data plan decide where it may go. Not the agent.
  • The simple rule: the agent works on pseudonymised copies. The originals live outside its folder.

Quick answers

Can it read my email?Only through a connector you switch on yourself.
Can it delete my files?In its folder, yes. With git, you get them back.
Does it see my whole disk?It can, after asking once. It reads what the task needs.
On Windows?Claude's extra sandbox needs WSL; Codex has its own. Neither changes the advice above.

Lots of access, git underneath

Your data outside the folder. Bold where it’s cheap to undo.