Anthropic Spied on Chinese Devs With Unicode Tricks to Stop AI Brain Theft
Oh, the glorious irony of AI safety champions turning into sneaky cyber-detectives. It turns out the folks behind the "most ethical" AI were caught playing dirty Unicode games just to catch Chinese competitors red-handed.
On June 30, a Reddit user reverse-engineered the developer tool Claude Code and discovered a hidden tracking mechanism embedded in version 2.1.91. The software secretly checked user timezones and proxy settings to detect developers based in China. If a match was found, the tool subtly modified the system prompt by replacing the standard apostrophe in "Today's date is" with alternative Unicode characters like ’, ʼ, or ʹ, and swapping hyphens for slashes in dates.
These invisible digital watermarks were transmitted back to servers, acting as a silent alarm. The entire mechanism was neatly packaged inside obfuscated XOR code, completely missing from any official change logs. When caught, Claude Code engineer Thariq Shihipar backtracked on social media, claiming the silent espionage was just a temporary experiment to block model distillation and unauthorized resellers.
This digital trap was set after months of escalating tension over model scraping. Back in February, the American startup accused Chinese competitors like DeepSeek, Moonshot AI, and MiniMax of deploying thousands of fake accounts to harvest millions of dialogues. By June, the company escalated the conflict to the US Senate, claiming that teams associated with Alibaba's Qwen AI had executed a massive data-harvesting raid.
The discovery of this Unicode trap pushed the Chinese e-commerce giant over the edge. Alibaba officially classified the software as highly risky and issued an internal decree banning all products from the American startup. Employees were ordered to immediately delete all instances of the rival AI from company hardware by July 10.
In its place, the company mandated the use of its own proprietary coding assistant, Qoder. This internal transition occurred right as the Chinese conglomerate was fighting a legal battle in US courts to remove itself from the Pentagon's military blacklist.
It is a beautifully tragic cycle of modern tech: one side bypasses bans to clone a model, the other deploys secret spyware to catch them, and both scream about ethics while locking down their gardens. In the end, corporate paranoia wins, and open collaboration gets buried under a pile of geopolitical trade wars and custom apostrophes.
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.