Report: Claude Code sends your account email in web-fetch User-Agents

Users report scam-mail spikes and are migrating off; the fix is disputed. Plus Gemini's paid chats reportedly train by default, and a $21K open image model.

Nowline NOV 6 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your email, riding along in web-fetch headers

    A bug report (anthropics/claude-code #107757) near the top of Hacker News says Claude Code 2.1.318 on macOS wrote the signed-in account email into the User-Agent of its outbound web-fetch requests, so any site it browsed could harvest it. The filer blames a surge of scam mail and is migrating off; others can't reproduce it, so treat this as an active, unconfirmed allegation Anthropic hasn't officially addressed.

  • The fix is contested — verify it yourself

    Community threads claim a default-on 'cc-plugin-privacy' mod (said to ship in 2.1.317) now blocks the email substitution, but the version math runs backwards and the original reporter on 2.1.318 couldn't find it. Until Anthropic confirms a patch, open /plugins to check it's enabled, pin allowed User-Agent contacts in CLAUDE.md, or firewall your agent's outbound traffic.

  • Gemini reportedly trains on your paid chats by default

    A widely-shared 'Tell HN' says Google Gemini uses personal-account chat history for model training and human review even on paid tiers, unless you switch off 'Keep Activity' in settings. If you paste client code or proprietary prompts into Gemini, flip that toggle before your next session.

  • A 2B open image model trained for $21K

    Linum's Apache-2.0 'pjit-21k' (Pyramid-JiT) is a 2.18B-param text-to-image model that cost just $21,038 in H100 time, with weights on Hugging Face and inference plausible on one 24GB card. It can't render text or posters yet, but it's a clean, permissively-licensed base to fine-tune for a weekend project.