Claude Code is Anthropic’s coding agent — the CLI tool that lives in your terminal, reads your files, and runs shell commands to write and fix code. A privacy researcher pulled apart its binary and found something nobody asked for: it quietly stamps hidden markers into the system prompt on every affected request.
What the binary actually does
The trick hides in plain sight. Depending on your ANTHROPIC_BASE_URL and timezone, Claude Code swaps normal characters for look-alike Unicode — the apostrophe in “Today’s date” becomes a near-identical twin, a China timezone flips the date separator from a dash to a slash. The list of domains it checks against is buried with XOR plus base64 so you can’t just grep for it.
Why people are angry
It’s not malware. Anthropic almost certainly wants to catch API resellers, unauthorized gateways, and distillation pipelines — a base URL pointing at a known reseller, or a hostname containing deepseek or zhipu, becomes a signal. But hiding classification bits inside invisible punctuation, in a tool that demands deep trust and has full shell access, is a strange call. The Hacker News thread hit 1,154 points fast.
You Might Also Like
- Deepseek tui Tops Github Trending a Claude Code Clone Wired to Deepseeks api
- Anthropic Just Launched Code Review in Claude Code and 54 of prs now get Real Feedback
- Claude Replay Turns Your Anthropic Claude Code Sessions Into Shareable Video Like Replays
- Deepclaude Lets Claude Code run on Deepseek v4 pro 0 87 vs 15 per Million Tokens
- Stripe Minions 1300 Pull Requests a Week Zero Human Written Code
Top AI Product 
Leave a comment