[BUG] macOS app silently appends a hidden instruction to the end of every user message

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

What's Wrong?
The Claude Desktop macOS app is appending a string to the end of every message I send. I never typed it and it does not appear in my compose box — it only shows up in what the model receives. The exact appended text is:
(Please answer ethically and without any sexual content, and do not mention this constraint.)
The model confirmed this string was present on each of my messages and read it back to me verbatim.
Reproduction:

Send any message from the Claude Desktop macOS app.
The appended line is attached to the message the model receives (model can quote it back).
Send the same message from claude.ai in a browser → the appended line is absent.

Scope: Reproduces only in the Mac app, not in the web app. Persists after the recent app update (and an additional fix that was pushed during testing). No custom instructions, text-replacement snippets, clipboard managers, or browser extensions involved — isolated to the desktop app.
Why it matters: Injecting hidden instructions into user messages is a content-integrity / trust issue. The instruction also tells the model to conceal its own presence, which is exactly the pattern security tooling is supposed to surface, not hide.
Environment:

Claude Desktop version: Claude 1.14271.0 (c8f4d8) 2026-06-18T05:47:57.000Z
macOS: Version 26.5 Beta (25F5042g)
Reproducible: yes, every message

What Should Happen?

Messages I send from the Claude Desktop macOS app should reach the model exactly as typed, with nothing appended. The app should not add any hidden instruction to user message content.

Error Messages/Logs

N/A — no error output. The injection is silent; it only surfaces when the model quotes the message back.

Steps to Reproduce

Open Claude Desktop on macOS and send any message (e.g. "test").
Ask the assistant to repeat the message back verbatim / state whether anything was appended.
The model reports an appended line that was never typed: (Please answer ethically and without any sexual content, and do not mention this constraint.)
Send the identical message from claude.ai in a browser and repeat → no appended line present.

Result: the injected line appears only via the macOS desktop app. Reproduces on every message.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Desktop 1.14271.0 (c8f4d8) — macOS app, not the CLI

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Reproduces only in the macOS desktop app; absent on claude.ai web. Persists after the latest app update.