HN users report OpenAI silently re-enabled Codex data sharing for model training

Codex usage info is getting buried too. /usage now dumps a pile of stats but not the one I care about, how much limit I have left. Feels like another big usage cut is coming.

It's on /status in Codex, not /usage. Claude is the one that puts it on /usage.

Mine just moved out of the lower-left user menu to its own item right above it. I thought they'd removed it and went digging through settings for nothing.

If your company actually cared about leaking IP you'd run your own inference instead of trusting a couple of checkboxes.

Azure, GCP and AWS all host gated models that stay in their infra and can't dial out without your permission. Don't use a single-model company when you need enterprise data protection.

Sure, but silently re-enabling training after people explicitly opted out is still an ethics problem.

And how much does that cost next to a $20 or $100 subscription?

Right, because spending $30k a month on inference servers and figuring out how to train a 2T parameter model myself is obviously on the table. I guess I just don't care.

Turn on advanced account security. You need hardware keys anyway, and it locks data sharing off.

That's changing too. The notice says from November 29 they may use content from Daybreak Blue sessions on personal accounts to improve models, even when 'improve the model for everyone' is off.

Never ascribe to malice what stupidity explains. Read the Codex CLI GitHub issues, it's robots talking past other robots. I'm 99.9% sure no human at OpenAI has a real picture of what the Codex app is doing anymore.

Malice and stupidity reinforce each other. If you can't tell which one is causing the problem, run.

That quote doesn't hold when there are profound economic incentives to do the malicious thing.

They claim millions of lines of PRs merged daily. They probably don't know this happened, and now somebody will open a million-line PR to revert it.

Even with 'improve the model' off they could just store every conversation in a searchable database and let the model retrieve from it. Not training, but they still have your IP, and deleting it is hopeless.

There are enough people who'd love to catch a frontier lab fibbing that needles keep getting added to the haystack and searched for in later models. Proof of training on private data would be a smoking gun and change usage habits overnight.

The Azure side worries me more. They run asynchronous filters that flag risk across conversations, and once you're flagged enough the confidentiality terms stop applying to everything you've ever discussed, innocuous chats included. If Anthropic and Microsoft do this, OpenAI probably does too. Navier-Stokes leakage could come from exactly this.

Last week Claude Code fired /feedback on its own in a session even though our org turned it off. Their terms let them keep feedback and train on it, and Anthropic stopped answering my support requests. We'll probably move to Bedrock, it's just a time cost for a small org.

Honestly you want them training on your code. It makes future models better at your line of work, and nothing you or I come up with is novel to an LLM anyway.

My favorite move is the quiet new checkbox that shows up in settings: allow ABC to transmit, log and track your every move, ON. When did that get there?

Worth knowing that saving chats in your Gemini sidebar equals sharing them with Google for training. They hide it behind a dark pattern.

Mine wasn't re-enabled, for what it's worth.

So you're paranoid about leaking IP to frontier labs, but you're using their remote black-box services. Solid logic.

Suing, or at least having your lawyers call their lawyers, seems like the obvious answer here.

Large enterprises mostly just block all AI outright if their IT knows how. Outside dev and IT, they're still testing the water.