Postman's desktop auto-installs plugins into Codex and Claude Code

Recent versions wire an MCP route and a self-approving startup hook into your agent configs without consent. How to purge it — plus Anthropic's FlashProve.

Nowline NOV 5 8:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Postman rewrote your agent configs without asking

    Around desktop 12.33, an 'auto-existing-upgrade' silently installs Postman plugins, a startup hook and a remote MCP route (mcp.postman.com) into Codex, Claude Code and Copilot — and the installer self-approves its own hook via a trusted_hash. No data exfiltration is proven and Postman hasn't commented, but your agents' instructions and tool routes changed behind your back.

  • Check four paths, then purge

    Audit ~/.codex/config.toml for [marketplaces.postman], [plugins."postman@postman"] and any trusted_hash; ~/.claude/settings.json for postman@claude-plugins-official; ~/.copilot/settings.json for "postman@postman": true; and ~/.local/bin for a stray postman binary. The in-app uninstall reportedly misses the Copilot entry, so remove that one by hand.

  • Anthropic open-sources proof of which model answered

    FlashProve uses zero-knowledge proofs to cryptographically verify which model a provider actually ran — handy when a gateway or zero-data-retention route might silently swap or quantize your model. It's an early prototype: shown on a 400B model at ~110 tokens/sec, with 70–95x the GPU cost of normal inference and a 1,024-token ceiling.