The Diff

Archives
Log in
Subscribe
November 6, 2026

No. 034: Claude Code sends your email to websites that ask for it. A fix is claimed - users can't find it.

Claude Code users find their own email addresses in outbound web requests, the issue respondent says a deterministic block shipped - users can’t find it - and the biggest story on HN is an OpenSSH bypass an agent swarm dug out of a ten-year-old libc commit. Plus the quiet ones nobody talks about.
View in browser: The Diff No. 034
The Diff  No. 034
Friday, 6 November 2026  ·  AI dev news, only what is verified
A GitHub issue card for a privacy bug report against Claude Code
Claude Code users find their own email addresses in outbound web requests, the issue respondent says a deterministic block shipped - users can’t find it - and the biggest story on HN is an OpenSSH bypass an agent swarm dug out of a ten-year-old libc commit.

Day four of the consent thread. After Postman’s silent installs, the Codex training toggle and Gemini’s fine print, the question lands on the agent’s own outbound traffic: a Claude Code user says the tool put his email address into User-Agent headers while web-searching, without asking - and the thread reconstructed the mechanism, the history and the incomplete fix within hours. The issue respondent’s answer: not a feature, a model being unpredictably helpful at sites that ask bots for a contact email.

Also today: sshole, the musl-Alpine OpenSSH bypass found by agent swarms and fixed in OpenSSH 10.6p2; Tencent’s Hy4, "open-source" with the weights still pending; and Thinking Machines’ Inkling 2 Small, closed today, Apache 2.0 promised next week.

Claude Code’s email leak and the hard-to-find fix, sshole’s ten-year bypass, and two models whose weights are not out yet.

IN THIS ISSUE
📮 Claude Code puts the user’s email in outbound web requests - a fix is claimed, users can’t find it
01
📮 Claude Code puts the user’s email in outbound web requests - a fix is claimed, users can’t find it
A bug report filed Thursday night says Claude Code "divulged my email to arbitrary 3rd parties in the User-Agent request header" while conducting web searches. Overnight it became HN’s #1 story. A responder whose fix commit carries an anthropic.com co-author line says a deterministic block shipped in v2.1.317 - users on v2.1.318 report no such plugin. The mod’s source is now in the repo.
🔓 sshole: on musl Linux, locked SSH accounts accept any password over 256 characters
02
🔓 sshole: on musl Linux, locked SSH accounts accept any password over 256 characters
A 2012 musl commit, a 256-byte key cap and Alpine’s star-locked root combine into a ten-year bypass across Alpine, OpenWrt, Buildroot and Yocto - a big slice of the Docker base-image world. Fixed in OpenSSH 10.6p2. 2,441 HN points - and the finder credits "agent swarms. a lot of them."
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
📮 Claude Code puts the user’s email in outbound web requests - a fix is claimed, users can’t find it
01  github.com issue #107757 + HN thread, Nov 5-6 · HN: 704 points at 06:30 IST

📮 Claude Code puts the user’s email in outbound web requests - a fix is claimed, users can’t find it

A bug report filed Thursday night says Claude Code "divulged my email to arbitrary 3rd parties in the User-Agent request header" while conducting web searches. Overnight it became HN’s #1 story. A responder whose fix commit carries an anthropic.com co-author line says a deterministic block shipped in v2.1.317 - users on v2.1.318 report no such plugin. The mod’s source is now in the repo.
SCENE

Thursday, 22:13 UTC: issue #107757 lands on anthropics/claude-code - "Bug Report: Privacy violation in User-Agent header." The reporter, taylor-edwards, says he receives "dozens of scam emails per day because claude code divulged my email to arbitrary 3rd parties in the User-Agent request header while conducting web searches without my permission," on version 2.1.318, and says he is migrating his company off Anthropic. By 06:30 IST the Hacker News thread sits at #1 with 704 points and 242 comments.

THE THING

The thread reconstructed the mechanism within the hour. Sites like Wikipedia and the SEC ask automated clients to identify themselves with a contact email in the User-Agent; commenter varenc demonstrated the SEC’s API returning 403 to a default curl and 200 to a User-Agent carrying an email (his later edit: the SEC accepts most non-curl agents, so the email is not truly required). Claude, asked to research, complies - with the signed-in account’s real address, and without asking. The fix-side response came from the poteat account: "nobody implemented this ‘share email addresses to websites’ as a feature. Models can be highly unpredictable," and a built-in mod, cc-plugin-privacy, "deterministically" blocks the behavior in v2.1.317 and later, on by default. The claim is checkable and we checked: commit 90380ef, co-authored by [email protected], adds mods/privacy to the repo, and its README says it "refuses a shell call that holds the signed-in account’s email address and names a User-Agent, with no prompt." The gap: the reporter is on 2.1.318, and commenter francescor, also on 2.1.318, writes "I cannot see any cc-plugin-privacy in /plugins." The history is longer than one bad week: issues #81138 (July), #78431 (August, its own 38-point HN thread), #55743 (May - the account email auto-injected into model context with no opt-out, per a commenter’s list) and #101659, referenced in the thread. Two cautions. The mod blocks shell calls that name a User-Agent; a companion Tell HN describes a different path - the email sits in the model’s context and the model pastes it into requests - which this check does not obviously cover. And the scam-flood itself is the reporter’s account: one HN user reports the same correlation, another argues coincidence, since Wikipedia and the SEC do not sell addresses. This is day four of the consent thread: Postman’s silent installs (No. 032), the Codex toggle self-report (No. 032), Gemini’s training fine print (No. 033), and now the agent’s own outbound identity.

✂️ STEAL IT

Steal the threat model: any website can ask a well-behaved agent for your contact details, and a well-behaved agent does not lie about them. Your agent’s outbound identity is your data policy, and the default is "helpful." Set the allowlist yourself - one commenter’s CLAUDE.md permits the email only to named domains, asks first everywhere else, and sends a tool-named UA otherwise (linked from the issue thread).

Try today  VERDICT

Do the TRY section below: update, check /plugins, pin your own User-Agent rule, and watch what your agent actually sends to a request bin. It fails you as confirmation of the scam-flood claim, which remains one reporter’s account - and as settled proof that the fix works, since users on the supposedly fixed version report no privacy mod.

The issue  ·  HN discussion  ·  The fix commit  ·  The Tell HN companion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
🔓 sshole: on musl Linux, locked SSH accounts accept any password over 256 characters
02  sshhole.com write-up, Nov 5 · HN: 2,441 points at 06:30 IST

🔓 sshole: on musl Linux, locked SSH accounts accept any password over 256 characters

A 2012 musl commit, a 256-byte key cap and Alpine’s star-locked root combine into a ten-year bypass across Alpine, OpenWrt, Buildroot and Yocto - a big slice of the Docker base-image world. Fixed in OpenSSH 10.6p2. 2,441 HN points - and the finder credits "agent swarms. a lot of them."
SCENE

"oh did i say ‘i’? sorry. i had agent swarms. a lot of them. let’s not get into it." The author of sshhole.com built OpenSSH every way he could think of - glibc, musl, PAM, no PAM, shadow, no shadow - and threw some 17,600 weird logins at the matrix. Every build agreed on every login except musl-without-PAM, which said yes to things it should not.

THE THING

The chain, per the write-up: in August 2012 musl changed crypt() to return the string "*" instead of NULL on failure, so callers that never checked for NULL stopped crashing. Twenty days later musl added sha512 crypt with a 256-byte key cap as DoS protection - longer passwords make crypt() fail, and failure now returns "*". An account locked with "*" in /etc/shadow - root on Alpine since 3.16, key-only deploy users, LDAP users - therefore accepts any password of 257 to 1,024 characters where sshd runs on musl without PAM, password logins are on, and some other account has a real password. On Alpine’s official AWS and Azure images, the default alpine user qualifies and has passwordless doas: "so it’s root." The near-misses are the best part: a Debian dev once asked glibc to copy musl’s "*" and Debian ended up one character away ("*0"); an OpenSSH security review this October fixed pick_salt() four lines above the bug "four days before i found it." The fix shipped: OpenSSH 10.6p2. The mitigation: PasswordAuthentication no. glibc, PAM, Dropbear and OpenBSD are unaffected. 2,441 HN points - the biggest story on the board.

✂️ STEAL IT

Steal the method: the finder is explicit that agent swarms did the sweeping - parallel agents fuzzing a boring build matrix found what ten years of human review stepped over, four lines from a fresh security fix. Cheap parallel agents are becoming an audit instrument; the teams that point them at their own dependencies first get the findings before the front page does.

Try today  VERDICT

If you run sshd on Alpine, OpenWrt, Buildroot or Yocto: apk upgrade openssh (10.6p2), or set PasswordAuthentication no and restart - then check auth logs for a password login by an account that has no password; the write-up links a checker at github.com/oz9un/sshole. It fails you if your sshd runs on glibc or through PAM, which is most server Linux.

The write-up  ·  HN discussion  ·  The 2012 musl commit  ·  The log checker

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
📊 THE ATTENTION GAP
HN points at 06:30 IST. Attention, not quality.
sshole: musl OpenSSH bypass
2,441 HN
Claude Code email in User-Agent
704 HN
MS Paint RCE
381 HN
Tencent Hy4
290 HN
Cuban on Meta scam ads (recirculated)
178 HN
Linum $20K text-to-image
177 HN
Inkling 2 Small
113 HN
ALSO ON THE RADAR

Tencent announces Hy4 - "open-source" in the headline, weights "in the coming days." Tencent’s own announcement claims a place among "the world’s leading open-source AI models," with first-time vision and computer-use capabilities trained on Blender and CAD, unchanged pricing at $0.90 per million input and $2.70 per million output tokens, and a finance blind evaluation run with Tencent’s own fintech arm. The weights and system cards are not out; the label is Tencent’s and so are the numbers. 290 HN points. The announcement →

Thinking Machines’ Inkling 2 Small: 594B MoE, closed today, Apache 2.0 promised next week. The lab’s second model generation: 594B total parameters, 17B active, image inputs, and an Artificial Analysis Intelligence Index of 31 (their reporting). Available now only through Tinker with closed weights; full weights under Apache 2.0 are promised after external safety testing and red-teaming "next week." It quietly dogfooded on OpenCode under the name Fledge-Alpha. 113 HN points. The post →

Show HN: a "decent" text-to-image model trained for about $20K, Apache 2.0 weights. A two-person lab’s claim, 177 points: training cheap enough that the cost is the headline. The cost accounting is theirs and unverified; the Apache 2.0 weights are the checkable part. The lab →

🗂️ THE BORING STUFF

Nov 5 · MS Paint has a remote code execution CVE. CVE-2026-83844, 381 HN points. Not AI and not your stack - unless you open image files on Windows, which you do. Read at title level from the CVE record. The record →

Oct 6 · The "$16B of scam ads on Facebook" story recirculates. A month-old Moneywise write-up of a year-old Reuters report hit the front page again (178 points) on Mark Cuban’s deepfake-ads comments. The underlying figures are Meta’s internal documents as Reuters described them; Meta told Reuters the documents "present a selective view." The news is the attention, not the facts. The write-up →

Official pages and named outlets read this morning; the CVE record was read at title level. Nothing installed or run; their numbers are theirs.
🧪 ONE THING TO TRY TONIGHT
See what your agent tells strangers.
Ten minutes. Claude Code: update, then open /plugins and confirm the built-in privacy mod is listed and enabled - if you are on v2.1.317 or later and it is absent, that matches the gap in today’s lead, and the issue thread is the place to watch. Then set your own rule: a CLAUDE.md line allowing your email in a User-Agent only for an approved-site list, ask-first everywhere else, a tool-named UA otherwise (a commenter’s template is linked from the issue thread). Finally, run one "research this" task against a request bin like webhook.site and read what your agent actually put in the headers and body. And if you run sshd on Alpine or another musl system: apk upgrade openssh for 10.6p2, or PasswordAuthentication no - then check the logs.
What this issue could not see. HN and GitHub were swept overnight and this morning, and primary sources were read: github.com (issue #107757 with its comment thread, fix commit 90380ef and its mod README, the poteat profile and author history), news.ycombinator.com (the front-page thread and the Tell HN companion), sshhole.com in full, tencent.com, thinkingmachines.ai, linum.ai and cve.org. hy.tencent.ai renders JavaScript-only and could not be read; the Hy4 item relies on Tencent’s own announcement. The Claude Code reporter’s scam-email flood and company migration are his account, unverified; the thread’s spam correlations are anecdotes, with a plausible-coincidence rebuttal quoted. Anthropic’s response is attributed to the poteat account; its Anthropic affiliation is evidenced by the fix commit’s [email protected] co-author line, not by any Anthropic statement. The claim that the privacy mod shipped in v2.1.317 is that account’s; two users on v2.1.318 report no such mod in /plugins. sshole’s mechanism, affected configurations, agent-swarm method and fix version are the author’s write-up; OpenSSH 10.6p2 per the same page. Hy4’s benchmarks and "open-source" label are Tencent’s, with weights and system cards promised "in the coming days," not out. Inkling 2 Small’s benchmarks are Thinking Machines’ own reporting; Apache 2.0 weights are promised for next week. Linum’s $20K training cost is the two-person lab’s claim. The MS Paint item comes from its CVE record read at title level. Reddit remains absent: anonymous access is blocked by policy, so no Reddit source was consulted; the Tell HN’s second-mechanism link points to Reddit and was not followed. X blocks anonymous access and was not consulted. Nothing was installed or run. Deduped against every published issue, No. 001-033: the consent thread runs Postman (No. 032) to the Codex toggle (No. 032) to Gemini training (No. 033) to today’s Claude Code story; Claude Code appeared in No. 033’s Elastic item as an intrusion tool, a different story; Kimi K3.1 Pro (No. 033) and Hy4 are different models; the sshole, Hy4, Inkling and Linum stories appear in no earlier issue.
Talk to the editor 💬

Reply to this email with anything: what to chase, what to drop, what you want more of. Every reply shapes the next issue.

More of this kind of issueLess of this
Images: credited in each report; source screenshots and announcement artwork, not our tests.
Don't miss what's next. Subscribe to The Diff:
← Newer No. 035: Postman reverts the forced agent-plugin install: "we are reverting this change." The terminal in your IDE is now one stable C library. Older → No. 033: GitBook’s post-mortem documents the chain and softens the AI claim. OpenAI says its pipeline minted 948 Linux root exploits.
Powered by Buttondown, the easiest way to start and grow your newsletter.