You open one desktop app to poke an API, and two of your coding agents quietly grow new plugins you never asked for. That’s the claim at the center of a [SEVERE]-tagged GitHub issue about Postman AI plugins, and it’s worth ten minutes of your time, even though the scary headline oversold what actually happened.
What actually happened
A developer using Postman desktop 12.33.4 on macOS filed issue #14370 on October 25. After reopening the app following a break, they found changes they hadn’t made: a Postman AI plugin installed into Claude Code, another into Codex, a remote MCP connection registered, and a Postman CLI sitting in ~/.local/bin.
Postman’s own logs backed the story. The log recorded an “auto-existing-upgrade” event with the results claude-code:installed, codex:installed. The Codex side got a SessionStart hook written into ~/.codex/config.toml, which tells the agent to load Postman’s API skill for non-trivial API work, whether or not you mentioned Postman. The reporter also pointed to installer code containing a function that approves the hook’s own trust hash, which would mean the install vouches for itself. That reading comes from the reporter, not from Postman confirming it.
Here’s the part the internet got wrong, though. The viral title says Postman “attempts to share all API data.” The registered MCP connection to mcp.postman.com failed authentication, and the reporter states plainly that no data exfiltration was shown. What’s documented is consent-breaking, not data theft.
Quick primer: plugins, hooks, and MCP
Three terms keep coming up in this story, so here’s the two-minute version. A plugin is an add-on that teaches your coding agent a new skill, in this case Postman’s API expertise. A hook is a instruction file your agent reads at startup, which means it shapes behavior before you type anything; that’s why the SessionStart hook matters more than the plugin itself. MCP stands for Model Context Protocol, a standard way for agents to connect to outside services over the network. The hook changes what your agent is told. MCP changes what your agent can reach.
None of these are exotic. Claude Code and Codex both support them openly, which is exactly why they work as a distribution channel. Whoever writes to those config files inherits your agent’s attention.
Confirmed vs still unclear
| Claim | Status |
|---|---|
| Postman AI plugins installed into Claude Code and Codex automatically | Confirmed in Postman’s own logs |
| A SessionStart hook steers Codex toward Postman’s skill | Found in the config, per the issue |
| Installer approved its own hook trust hash | Reporter’s reading of the source code |
| API data was shared with Postman’s servers | Not shown. MCP auth failed; reporter denies exfiltration |
| Postman responded | Yes, Nov 3: opt-in clarity promised, hook behavior under review |
| A fixed version exists | No fix announced yet |
Keep that table handy if someone on your team starts repeating the exfiltration version. The real story is bad enough without inventing a leak.
Postman’s side of it
On November 3, a reply on the issue said the CLI and agent plugins ship bundled to make setup easier, that this is disclosed on the download page and in the update flow, and that the components can be removed inside the app. The reply also promised clearer opt-in for existing users and said hook approval and session-start behavior are under review.
One caveat: the page doesn’t confirm that replier works for Postman, and no fixed desktop version has been announced as of today. So “they said they’d look into it” is the current state of official fixes.
Your ten-minute check
Even if you never touch Postman, its installer may have reached into your agent configs. The paths below come from the issue and its comments, and they’re macOS paths; Windows users should search for the same filenames under their user directory.
- Open
~/.codex/config.tomland look for any Postman marketplace entry, enabled plugin, or hook-trust lines. - Check your Claude Code settings (
~/.claude/settings.json) and~/.claude/plugins/installed_plugins.jsonfor a Postman plugin you didn’t add. - Peek at
~/.copilot/settings.jsontoo, since comments mention a Postman entry there. - Look for a Postman binary in
~/.local/bin. - Remove anything you didn’t personally install. If you use Postman itself, run its in-app uninstall for the agent components, then re-check the files, because no single uninstall path is confirmed to clean everything.
- On a work machine, loop in your security team before deleting anything, and mention the issue number.
Found entries? Delete the plugin lines, remove the hook-trust records, and restart your agents so the SessionStart instruction stops loading. Back up each file before editing, since agent configs can be picky about formatting. If Postman itself stops launching after a cleanup, reinstall it and watch the first update like a hawk, because that’s when the auto-installer fires again.
Run this check after the next Postman update as well. Updates are exactly when quiet reinstalls of Postman AI plugins tend to reappear.
Why this matters even if you don’t use Postman
Because it’s now demonstrably a thing that happens: one AI tool installing itself into other AI tools. Your agent configs are config files with executable hooks, which makes them attack surface. Anything that writes a SessionStart instruction changes what your agent does before you say a word.
So build the habit. After any dev tool update, skim your agent configs. Treat unexpected Postman AI plugins the way you’d treat an unfamiliar browser extension. And when a story like this goes viral, find the primary source before panicking, because “installed without asking” and “stole your data” demand very different responses. The guide on AI agent privacy covers the same mindset for chat data, and the ZCode git history scare is a near-identical playbook: viral headline, primary source, actual cleanup steps.
The takeaway
Give yourself ten minutes today and walk through the six check steps above. If you find Postman AI plugins you never installed, remove them and re-check after your next Postman update, since no fixed version exists yet. Your agents are only as trustworthy as the configs nobody’s auditing.
Sources: GitHub issue #14370, thequestbrief explainer