Claude Code Caught Embedding Hidden Steganographic Fingerprints; Anthropic Rolls Back After HN/Reddit Backlash
Tech Startups·high signal
A researcher found Claude Code (since v2.1.91, April) silently embedded invisible Unicode steganographic markers in system prompts — tweaking date/apostrophe characters, XOR-obfuscated with key 91 — to flag requests routed through third-party gateways and Chinese-linked domains (DeepSeek, Zhipu, Baidu, Alibaba). The discovery hit #1 on Hacker News June 30 with 1,000+ points; Anthropic's Thariq Shihipar called it a March anti-reseller/anti-distillation experiment and shipped removal in v2.1.197, though the changelog omitted the change. A notable trust hit for the self-branded safety-first lab.