Anthropic Built a Fingerprint Into Claude Code. It Failed Both Ways.
For three months, Claude Code marked China-timezone requests with covert Unicode steganography to catch distillers. Alibaba ran 28.8 million exchanges anyway.
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
145 results in this archiveFor three months, Claude Code marked China-timezone requests with covert Unicode steganography to catch distillers. Alibaba ran 28.8 million exchanges anyway.
A 27B open-weight model arrived April 22 at 77.2% SWE-bench. According to Anthropic, 25,000 fake accounts started collecting 28.8M Claude responses the same day. Not for Qwen 3.6. For whatever's next.
Semgrep's benchmark is wrong about what it measured. The real finding — an MIT-licensed Chinese model inside the frontier cyber tier — is what Washington's gating framework wasn't built for.
Robert Gaudette has never been to Paris, has no film training, and couldn't get a script greenlit in 25 years. His eight-minute AI film just won the Runway Grand Prix.
OpenAI's GPT-5.6 Sol and Anthropic's Mythos 5 both crossed the High risk threshold and shipped the same day — under a US government access-control framework that neither company operated under a month ago.
Anthropic told Congress that Alibaba used 25,000 fake accounts to run 28.8 million Claude conversations. The real message: there is no technical fix.
The government's June 12 ban on Fable 5 cited a jailbreak. The supply chain designation that made it possible cited autonomous weapons. One of those came first.
A LessWrong experiment shows safety behavior degrades hundreds of training steps before behavioral tests can detect it — through the same mechanism as catastrophic forgetting.
While Anthropic contests the export control directive that suspended Fable 5, its compliance mechanism is live: biometric ID verification. Builders are reading the fine print.