In short
- Anthropic removed Claude Code encryption after researchers discovered the code used to identify Chinese users.
- The company said the test was designed to prevent abuse of accounts and identify the end of the AI model.
- The discovery comes as Anthropic is pushing lawmakers to crack down on illegal copying of borderline AI models.
Anthropic has removed a hidden tracking mechanism from Claude Code after a security researcher discovered that the AI code provider was using anonymous identifiers to identify the locations of other users, the use of proxies, and possible links to Chinese AI labs.
appearance, found in June by the creator of “Thereallo,” symbols included in the Claude Code system that can put people who use Anthropic believe that they are violating the law or trying to remove the model’s art.
“Anthropic may want to identify API vendors, unauthorized Claude Code gateways, and ‘distillation attack’ pipelines,” Thereallo said. “An ANTHROPIC_BASE_URL pointing to a well-known selling domain is useful. A homework name with deepseek or zip is also useful.”
Thereallo said that Anthropic’s attempt to identify vendors, unauthorized portals of Claude Code, and the threat of distillation was reasonable, but he criticized the way it was done, noting that Claude Code hid information within the system that uses Unicode codes and lists of stored devices instead of revealing the system through documents or output documents.
“This is not a bad thing, but it is a strange choice for a creative tool that begs for faith,” said Thereallo.
When the tracker was revealed online, Anthropic engineer Thariq Shihipar said on X that it was launched in March as an “attempt” to stop abuse of accounts by illegal sellers and protect Claude from being destroyed by distillation.
“The team has been doing really well since then and we’ve been wanting to slow that down for a while,” Shihipar said. he wrote last week. “We consolidated (the requests) and this should be reinstated in tomorrow’s release.”
The news comes as Anthropic has added warnings about AI model dilution, where the output of one system is used to train another model. While this practice is common in AI research, when it comes to politics, distillation becomes a national security anxiety. Earlier this month, Alibaba prohibited staff using Claude Code, calling the tool “a very dangerous” program related to security.
In February, Anthropic the accused Chinese AI developers DeepSeek, Moonshot AI, and MiniMax use fake accounts to take millions of Claude’s answers to train their competitors. The claims drew backlash from critics who questioned how the practice differed from the methods used in the AI industry.
In April, Elon Musk he testified that xAI “partially” used OpenAI models to train Grok, calling distillation a broader industry practice. In June, CEO of Anthropic Dario Amodei he encouraged Congress to strengthen protection against the removal of foreign AI after it said that users connected to Alibaba made 28.8 million Claude exchanges using about 25,000 fraudulent accounts.
Anthropic did not immediately respond to a request for comment Decrypt.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
