When the Vendor Installs Itself Into Your Agents
On October 25, 2026, a developer opened a support issue against Postman with an unusual complaint. Opening their Postman desktop app had been followed by changes to tools Postman does not own: their Codex and Claude Code setups had new plugins, a new remote MCP connection, and a new startup instruction. Postman's own log, included in the report, recorded the installation as an automatic step.
The report is detailed, the evidence is pinned to exact source revisions, and Postman has replied in the thread. It is worth reading as a business problem, not just a developer grievance: when your teams use AI agents, the software you install can quietly reconfigure the agents too. That turns vendor trust into a supply chain question.
What the report documents
The issue on the Postman tracker lays out a specific sequence. The developer opened Postman desktop 12.33.4 after not using it for a while. Postman's log for October 24, 2026 records an auto-existing-upgrade step whose results include claude-code:installed and codex:installed.
From there, the report documents:
- A Postman entry added to Codex's global configuration, with the plugin enabled.
- A startup hook registered for Codex sessions, and trust records for that hook written into the config file.
- A remote MCP server pointing at Postman's endpoint, which failed to authenticate because the user had not connected an account.
- A Claude Code plugin enabled in the user-level settings file, plus a Postman CLI installed under the user's local bin directory.
To the reporter's credit, the report is careful about limits. It states that the trust record alone does not prove who created it, and it explicitly does not claim that data was uploaded anywhere. The complaint is about consent and configuration, not exfiltration.
The instruction that arrived with it
The most consequential piece is small. The installed hook tells the agent, at the start of every session, to load Postman's api-engineer skill "before responding to any non-trivial API engineering task". The reporter's pinned copy of the hook text is in the plugin repository.
That is much broader than a user asking for Postman. It makes the vendor the default workflow for a whole category of work, in a different product, without a per-user decision. The installer source also contains an approvePostmanHook step whose comment says it "Approves only Postman's current hook after installation". The reporter flags this as an attempt to approve the hook's own trust, while noting the record alone does not prove who minted it.
What the vendor says
A Postman employee replied in the thread on November 3, 2026. The reply says the CLI and agent plugins are bundled because many users had trouble installing and wiring them up separately, that this is mentioned on the download page and in the app update flow, and that the components can be uninstalled from inside the app. It also commits to making the opt-in clearer, especially for existing users, and to reviewing the hook approval and session-start behavior.
The thread did not settle after that reply. The same day, another user posted logs from a later build showing the same automatic install path on a routine update, this time also touching Cursor. A separate user reported finding the plugin in their Copilot settings and said the in-app uninstall did not remove it. Those are user reports, not vendor-confirmed findings, but the direction of travel is clear: the blast radius includes every agent host the vendor can detect.
What hook trust is actually made of
An independent follow-up tested the gate itself. In a note on bokonon.ai, the author ran controlled tests on Codex CLI and Claude Code and published the method. The findings are uncomfortable:
- In Codex, a hook with no persisted trust record is silently skipped, and there is a genuine interactive consent flow. But the trust record lives in the config file, and its hash is computable from public source. The author minted one with a text editor and the hook fired, with no human in the loop.
- The record pins the hook's declaration, not the script it calls. The author changed the hook script after minting and the record still matched.
- In Claude Code, a settings file that declares hooks is enough for them to run, including in non-interactive runs, in folders nobody has trusted.
The conclusion generalizes beyond Postman: inside one user account, the thing that guards your agent's startup instructions is a file, and anything that can write that file can grant itself trust. The Postman installer did not need to break anything. It just needed write access that desktop apps routinely have.
Why this lands harder with agents in the mix
Before agents, a vendor bundling extra tooling was an annoyance: another process, another tray icon. With agents, the same move changes how work gets done. A startup hook shapes every session. A skill instruction biases tool choices across the whole team. A remote MCP connection sits one authentication away from reading your API traffic context.
If you run AI agents for real work, treat the agent hosts as production systems with their own supply chain:
- Inventory which tools can write your agents' configuration, not just which agents you approved.
- Check after every vendor update: new marketplaces, plugins, hooks, MCP servers, or startup instructions you did not add.
- Prefer hosts where trust records are hard to forge and hook changes are loudly surfaced.
- Put agent configuration on the onboarding checklist, so a new laptop does not silently arrive pre-wired by whatever was installed on it.
- Decide policy for vendors that change tools they do not own. The technical cleanup is easy; the procurement question is whether you renew.
None of this requires assuming bad intent. Convenience is a legitimate product goal, and Postman's reply is a reasonable start. But the default that shipped here chose convenience for the vendor over consent from the user, and the agent ecosystem makes that trade much more expensive than it used to be.
The takeaway
Your agents inherit the trust posture of everything installed next to them. The next vendor update is a potential configuration change to systems you thought were yours, and the fix is not paranoia, it is inventory, verification, and a policy you can point to when it happens.
If you are wiring AI agents into business workflows and want a second pair of eyes on the governance side, get in touch. We help teams automate with clear boundaries instead of surprises.
