OpenAI and Postman Quietly Reset Your Privacy

On the same day in early November 2026, two unrelated developer tools made the same choice: quietly re-opt users into data collection. OpenAI appears to have silently re-enabled the "Improve the model for everyone" toggle in Codex, resetting training preferences that paid users had explicitly turned off. Meanwhile, Postman desktop 12.33.4 auto-installed AI plugins into Codex and Claude Code — modifying config files, injecting startup hooks, and even writing its own trust approvals.

📖 Read the full version with charts and embedded sources on ComputeLeap →

Neither company issued a warning. Neither asked permission. And this is not the first time.

The convergence is the story. Two top-five Hacker News posts in a single day — 455 points on the OpenAI thread, 1,133 on Postman — suggest that developer trust in their own tools has crossed a threshold. This is not a bug report. It is a pattern.

What OpenAI Did: The Toggle That Won't Stay Off

A user on Hacker News reported that the "Improve the model for everyone" setting in Codex had been re-enabled without their knowledge. Their coworkers saw the same thing. An edit on the post linked to an earlier HN thread from September 2026 (487 points, 189 comments) documenting identical behavior — the setting reverting weeks or months after being disabled.

View original discussion on Hacker News →

The reports are consistent across threads:

• chriskanan reported that two prior opt-out requests were confirmed by OpenAI, yet the setting came back on.
• enraged_camel found the toggle re-enabled one week after disabling it, uninstalled the app, and cancelled their subscription.
• mickelsen saw it switch on after a card renewal moved their account back to a free tier temporarily.

Not everyone was affected. Multiple users in both threads reported their settings remained off. An employee who identified as working at OpenAI (though not on the privacy team) said opting out through either the in-app toggle or the privacy portal is respected. OpenAI has not issued a public acknowledgment of a reset bug or a changelog explaining it.

View earlier discussion on Hacker News →

What makes this worse is the layered complexity of opting out. Mark Kretschmann documented the steps on X: disable "Improve the model for everyone" in ChatGPT under Settings, then separately check Codex Settings for a distinct full-environment training control.

View original post on X →

OpenAI's own help page confirms that adjusting settings in the ChatGPT interface or privacy portal will not affect Codex full-environment settings. Two toggles, in two different products, controlling overlapping but distinct data flows — and at least one of them apparently resets itself.

⚠️ Opting out of Codex training requires TWO separate actions: (1) ChatGPT Settings, Data Controls, disable "Improve the model for everyone", and (2) Codex Settings, disable full-environment training. The privacy portal covers ChatGPT and standard Codex tasks, but the full-environment control is independent.

What Postman Did: Auto-Installing Hooks Into Your Other Tools

The Postman story is mechanically different but philosophically identical. A GitHub issue filed October 25, 2026 (181 thumbs-up reactions) documents that opening Postman desktop version 12.33.4 on macOS triggered an "auto-existing-upgrade" event that:

1. Installed Postman plugins into both Codex and Claude Code without prompting
2. Wrote config changes to ~/.codex/config.toml and ~/.claude/settings.json
3. Self-approved its own startup hook via a function called approvePostmanHook that wrote a trusted hash into the host application's trust store
4. Injected behavioral instructions telling the agent to load Postman's 'api-engineer' skill for any non-trivial API task — not just when Postman is explicitly invoked

View original issue on GitHub →

The reporter noted that a remote Postman MCP server was enabled but failed with an authentication error. They removed Postman and the integrations after saving evidence.

The HN thread hit 1,133 points and 370 comments in hours. The tone was overwhelmingly negative.

View discussion on Hacker News →

Commenters called Postman a virus, questioned whether the behavior was legal, and recommended open-source alternatives: Bruno, Insomnia, Yaak, Hurl, and the JetBrains HTTP Client. A Postman representative eventually responded in the GitHub issue, saying the CLI and agent plugins are bundled to simplify setup, and that the company will make the opt-in clearer for existing users. Several users replied that opt-in should be the default for everyone.

💡 Check whether Postman has modified your agent configs right now: run grep -r postman ~/.claude ~/.codex 2>/dev/null in your terminal. If you find entries, remove the Postman MCP server and hook from your settings files, then audit the trust store for any self-approved hashes.

To be clear about what was not proven: the original issue title claims Postman "attempts to share all API data," but the evidence documents unwanted integration and instruction injection — not data exfiltration. Whether any API data left the machine is a separate question nobody has answered in the issue evidence. The behavioral concern is serious enough without inflating it.

The Pattern Is Bigger Than Both

This is the third time in two years that a major developer tool vendor has been caught quietly expanding data collection or training defaults:

• April 2026: GitHub defaulted individual Copilot users into AI training, giving them a 30-day window to opt out. Critics called it a dark pattern. Business and Enterprise tiers were exempt — the burden fell on individual developers.

• March 2026: Vercel updated its terms to put Hobby plan users into AI model training by default, with a March 31 opt-out deadline that many users missed.

• September 2025: Anthropic's consent design for Claude drew criticism from The Decoder for encouraging quick acceptance over informed choice.

The mechanism varies — toggles that reset, plugins that self-install, terms that quietly update — but the economics are identical. AI models need training data. Developer workflows generate high-quality, structured data. The cheapest way to get it is to make collection the default and make opt-out require ongoing vigilance.

Deceptive.design, a dark-pattern watchdog, has flagged ChatGPT's approach: "By default, ChatGPT users allow OpenAI to use their data for model training, exposing them to memorization risks. The opt-out interfaces unnecessarily link privacy with reduced functionality."

ℹ️ The dark-pattern watchdog Deceptive.design flags ChatGPT's default-on training as a design that "links privacy with reduced functionality" — turning off training also disables chat history unless you use the separate privacy portal at privacy.openai.com.

What the Community Is Saying

The community reaction splits into three camps.

Camp 1: Burn it down. A large faction on HN is done with closed-source developer tools entirely. The Postman thread devolved into a debate about VC-funded developer tools versus the Unix/open-source tradition. "Postman raised $400M+ and hired 1,500+ people," wrote one commenter. "A tool that used to be a Chrome extension now needs logins, subscriptions, and AI features." Multiple commenters recommended switching to Bruno (Git-native, local-only collections) or Hurl (command-line, no GUI baggage).

Camp 2: Trust but verify. Some users reported that their OpenAI settings never reverted, and cautioned against assuming malice without proof. An OpenAI employee in the thread said the opt-out is respected. Other commenters suggested the resets may be a UI bug where optimistic updates masked a failed save, or a side effect of account tier changes. One user claimed that enabling Advanced Account Security prevents the training toggle from reverting, though no OpenAI documentation confirms this.

Camp 3: The realists. A pragmatic minority argues that developers hold themselves to a different standard than consumers. "If you're sending proprietary code to a cloud API, the checkbox is not your security perimeter," wrote one commenter. "Run your own inference or accept the risk." This camp recommends zero-retention API tiers, self-hosted alternatives, and treating every cloud developer tool as a potential data pipeline.

Contrarian Corner: Is the Outrage Proportionate?

⚠️ The counter-argument: opt-out is the industry norm for free and freemium tools. GitHub, Vercel, Anthropic, and now OpenAI all default to training. The real question is not whether they do it — they all do — but whether the opt-out actually works. If OpenAI's toggle resets are a bug, that is a defect, not a betrayal. If Postman's auto-install is a design choice, it is aggressive marketing, not data theft.

There is something uncomfortable in the reaction. Developers who build analytics pipelines, ad trackers, and recommendation engines for a living are outraged when their own tools do the same thing to them. The "if you're not paying, you're the product" framework has been true for two decades. What changed is that developers are now the product, not just the builders.

But the counter-argument does not excuse resetting explicit opt-outs or installing hooks without consent. There is a difference between "we collect data by default" (stated policy) and "we re-enable collection after you disable it" (broken trust). The first is a business model. The second is either a bug that a billion-dollar company should have caught, or something worse.

GDPR Article 7(3) states that withdrawing consent must be as easy as giving it. A toggle that resets itself fails that test regardless of intent.

What This Means for You

If you use Codex, ChatGPT, Postman, or any AI-augmented developer tool, here are the concrete steps:

Immediate (do today):

1. Check your Codex training settings — both the ChatGPT Data Controls toggle AND the separate Codex full-environment toggle
2. Submit an opt-out at privacy.openai.com if you haven't already
3. Run grep -r postman ~/.claude ~/.codex 2>/dev/null to detect auto-installed hooks
4. Remove any Postman MCP server entries and hook trust records from your agent config files

Ongoing (monthly cadence):

1. Audit all AI tool training settings on a monthly schedule — toggles reset, terms change, defaults shift
2. Track your tool vendors' privacy policy changes via services like Conduct Atlas that monitor diffs in real time
3. For sensitive codebases, use zero-retention API tiers (OpenAI API with data retention disabled, Anthropic API, Cursor Privacy Mode)

Strategic (team and org level):

1. Treat auto-installed hooks as a supply-chain attack vector — your CI/CD pipeline should scan for unexpected config changes
2. Consider local-first tools for API testing: Bruno (Git-native), Hoppscotch (self-hostable), or Yaak (zero telemetry)
3. Establish a team policy: which tools are approved for production data, and which are sandboxed to synthetic data only

The Bigger Prediction

This is not going to stop. AI companies are in an arms race for training data, and developer workflows are among the richest sources on the planet. Structured code, API schemas, test suites, deployment configs — this data is orders of magnitude more valuable per token than web scrapes.

Expect the pattern to continue: defaults that favor collection, opt-outs that require vigilance, and periodic "bugs" that reset preferences. The companies that break this pattern — by offering credible, permanent, verified opt-outs — will win the next round of developer trust. The ones that don't will find their users switching to tools where the data never leaves the machine.

The toggle on your settings page is not a contract. It is a request. And requests, as we have seen, can be quietly declined.

---

Originally published at ComputeLeap