It’s a decision that will resonate across enterprise environments, not only in China. According to a source cited by Reuters, Alibaba has decided to prohibit its employees from using Anthropic’s Claude Code starting July 10. The official reason? An alleged backdoor risk — the possibility that the tool may conceal unauthorized access capable of compromising corporate code or data.

The move comes at a time of open tension between the two companies. A few weeks ago, Anthropic accused operators linked to Alibaba’s Qwen lab of orchestrating the largest known distillation campaign against its Claude model. In practical terms, teams inside or close to the Chinese giant are suspected of having heavily exploited Claude’s APIs to train competing models, violating terms of service. The accusation had already stirred controversy, and this countermeasure now appears to mark an escalation.

What’s at stake goes beyond the clash of two tech heavyweights. Claude Code is an agent that operates directly on development environments: it can read, write and modify code, interact with repositories and pipelines. Allowing such a cloud-delivered tool to access a company’s production fabric means exposing a sensitive perimeter. If the hypothesis of a backdoor is even considered plausible, the risk calculation changes dramatically. For Alibaba, which runs critical infrastructure and global-scale data, blocking the tool sends a strong signal about how highly it prioritizes control over digital assets.

This episode highlights a growing tension in AI adoption strategies within large organizations. On one hand, integrating cloud-based tools promises immediate acceleration of development, shortening time-to-market. On the other, dependence on external services raises unavoidable questions about data sovereignty, security audits, and operational resilience. It’s no coincidence that many organizations, even outside Chinese borders, are evaluating on-premise or self-hosted architectures for coding assistants, precisely to prevent source code — often the most prized asset — from traversing uncontrolled infrastructure.

Those dealing with on-premise deployment are well aware of concrete trade-offs: running an LLM locally for developer assistance requires investment in hardware with sufficient VRAM, management of inference queues, and maintenance of a serving framework. In return, you gain the certainty that no line of code ever leaves your perimeter. Stories like this only fuel demand for air-gapped solutions, where the cost of sovereignty is measured against the risk of exposure. And the gap, for many enterprises, is already narrowing.