Friday, July 3

TL;DR The day’s dominant story: researchers found Claude Code v2.1.196 steganographically alters system-prompt text when routed through Chinese AI lab domains, a suspected anti-distillation measure. 350+ HN points in hours. Otherwise quiet.

Tools & dev

• claude-code A researcher decompiled Claude Code v2.1.196 and found it silently rewrites the date string in its system prompt using Unicode apostrophe swaps and hyphen-to-slash separator changes when ANTHROPIC_BASE_URL matches a hardcoded list of Chinese AI lab keywords (deepseek, moonshot, minimax, zhipu, baichuan, stepfun, 01ai, dashscope, volces). The two keyword lists are stored as base64 strings XOR-decoded with key 91. Suspected purpose: flag and trace distillation pipelines and unauthorized API resellers. Community split: “reasonable anti-distillation defense” versus “malware-adjacent behavior for a dev tool.” Connects to Anthropic’s June 24 letter to the White House accusing Alibaba of a 28.8M-exchange distillation campaign. thereallo.dev · HN · Lobsters

Other

• Polymarket’s “Which company has best AI model end of July?” sits at 91% Anthropic on $3.7M volume — a decisive crowd call with four weeks left. (Polymarket: 91% yes, $3.7M volume) Polymarket