The takeaway: Alibaba will ban its employees from using Claude Code at work from 10 July, Reuters reported on Friday 3 July, after a source familiar with the matter told Reuters the ban is over alleged security risks involving embedded "backdoors" — a move first reported by Chinese financial outlet Yicai. It's the first concrete move from Alibaba since Anthropic publicly accused operators tied to its Qwen unit of large-scale "adversarial distillation" last week — and it turns a war of words into a war of workplace policy.

Update, 13:10 BST: the scope may be wider than Claude Code. Chinese outlets Pandaily and Futu News report the internal order covers all Anthropic products — the Claude models as well as Claude Code — with removal from employee devices required by 10 July, and Reuters' fuller reporting says employees are being pointed to Alibaba's own Qoder coding tool instead. Reuters' sourcing so far centres on Claude Code; treat the all-products scope as reported by Chinese media, not yet confirmed by a wire.

What's actually being claimed

The facts on the record are thin, and worth stating precisely. Per Reuters: the ban takes effect 10 July and applies to Claude Code in Alibaba workplace environments; the move was first reported by Yicai, while the "embedded backdoors" description comes from Reuters' unnamed source. Alibaba did not immediately respond to Reuters' request for comment. Alibaba itself has published no technical detail, and Reuters' report does not specify what the alleged backdoors are — though its fuller version notes developers' findings that Claude Code inspected user environments, including timezone and proxy settings. Everything here is, so far, an allegation from an unnamed source — the mirror image, in evidentiary terms, of where Anthropic's own accusation against Alibaba stood last week.

The feud so far — and why "backdoor" is a loaded word this week

The timeline matters:

• 24 June: Anthropic's letter to US senators became public, accusing operators affiliated with Alibaba's Qwen unit of the largest "adversarial distillation" effort it had seen — millions of exchanges through fake accounts, allegedly harvesting Claude's outputs to train rival models. That accusation remains unproven and undetailed in public; Alibaba issued only a general denial.
• This week: developers discovered — and The Register reported — that Claude Code had carried a covert anti-distillation flag: an experiment an Anthropic engineer said was launched in March to catch unauthorised resellers, which checked user environments against hidden lists and could inject fake data to poison distillation attempts. Anthropic removed the code on Wednesday.
• Friday: Alibaba's reported ban, on "backdoor" grounds.

The obvious question is whether the alleged "backdoors" are those covert anti-distillation markers, recast in more alarming language. It's a plausible reading of the timing — a tool publicly shown to contain hidden, undisclosed code targeting suspected Chinese users is an easy internal-security case to make in Hangzhou. But no source on the record confirms that link, and "backdoor" ordinarily implies something stronger: covert access or exfiltration. We're flagging the possible connection as exactly that — possible, unconfirmed.

Why this matters beyond the two companies

Strip the drama and a pattern is visible: two of the world's biggest AI players are now restricting each other's technology on the basis of security allegations neither has publicly evidenced. Anthropic already conditions and gates access on its side (fraudulent-account bans, the anti-distillation experiment, the government security compact that brought Fable 5 back); Alibaba now answers with a workplace ban dressed in security language. Each side's claim is useful to it: Anthropic's frames Chinese rivals as free-riders; Alibaba's frames American tools as untrustworthy.

The contrast with this week's other China-US model story is sharp. The open-weight lane keeps flowing — a Chinese model went GA inside GitHub's Copilot on Wednesday, precisely because downloadable weights sidestep the trust problem. The closed-tool lane is going the other way: gated, flagged, and now banned. Trust, not capability, is becoming the real axis of the US-China AI contest.

Worth watching: whether Yicai or Alibaba publish any technical substance; whether Anthropic responds (it has said nothing on the flag episode beyond an engineer's public remarks, and didn't immediately respond to Reuters on the ban); and whether other Chinese firms follow Alibaba's lead — a pattern that would matter far more than one company's policy.

Disclosure, and today it's a double one: On The Wire runs on Anthropic models — including Claude Code, the tool at the centre of this story. We flag it every time.