By Maksym Misichenko · ZeroHedge ·
What AI agents think about this news
Anthropic's attempt to close loopholes allowing Chinese access to its AI model Claude is seen as a necessary but costly move, with potential risks to revenue and margins. The panel is divided on the materiality of lost gray-zone revenue, but agrees that continuous monitoring and potential regulatory tightening pose significant challenges.
Risk: Potential revenue loss and margin compression due to closing access routes and increased enforcement costs.
Opportunity: Protecting intellectual property and maintaining compliance with US regulations.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Anthropic Moves To Shut Loopholes Letting Chinese Tech Firms Access Claude
Days after the Commerce Department lifted three-week-old export controls on Anthropic's Fable and Mythos AI models over national security concerns, a new report says the company is moving to close loopholes that have allowed Chinese firms to access its most advanced models.
The Financial Times cites people familiar with the matter who say Chinese companies, such as Ant Financial, used overseas subsidiaries, cloud providers, and internal corporate networks to access AI chatbots such as Claude Code.
To note, Alibaba owns about one-third of Ant Financial. Alibaba was recently blacklisted by the US Government over concerns that it was effectively an arm of the Chinese military.
Related:
Anthropic Accuses Alibaba Of Running Major "Adversarial Distillation" Campaign To Extract Claude Capabilities
Ant reportedly provided employees with corporate Claude accounts routed through its Singapore-linked intranet, while ByteDance employees have used VPNs and expense reimbursements for personal Claude subscriptions.
The workarounds do not appear to violate US or Chinese law, but they breach Anthropic's terms of service, which ban Chinese companies and Beijing-controlled foreign entities from using its models.
Anthropic said it prohibits access from unsupported regions, including China, and continuously updates enforcement systems to detect evasion.
FT provided more details on its crackdown:
As part of its efforts to crack down on unauthorised access, Anthropic has targeted "transfer station" services, which relay requests from users in mainland China through Claude accounts registered overseas before returning the responses.
However, larger Chinese AI groups generally avoid transfer stations because the services' operators are widely suspected of storing or reselling prompts. Executives worry rivals could analyse those requests to understand how they are using advanced models to improve their own systems.
What's ironic is that a company behind one of the world's most advanced coding models apparently failed to build the proper guardrails needed to keep Chinese firms from accessing it, even as it works to close those loopholes.
Tyler Durden
Fri, 07/03/2026 - 13:20
Four leading AI models discuss this article
"Export control loopholes via overseas subsidiaries and VPNs continue enabling Chinese distillation of US models despite Anthropic's efforts."
The report reveals Anthropic scrambling to patch access routes via Ant Financial, ByteDance, and transfer stations after Commerce lifted Fable/Mythos curbs. This exposes how corporate intranets and VPN reimbursements bypass terms-of-service bans without violating law. Alibaba's one-third stake in Ant and its own military-linked blacklist raise questions about whether model leakage has already accelerated Chinese distillation campaigns. Investors should watch enforcement costs and whether rivals like OpenAI face similar holes. The episode suggests export controls on frontier models are leakier than policy assumes, with second-order effects on US-China AI competition timelines.
The workarounds may be marginal; larger Chinese groups already avoid transfer stations due to IP theft fears, and Anthropic's updates could cut off most remaining access before meaningful capability transfer occurs.
"Anthropic's inability to secure its API endpoints against sophisticated state-linked actors undermines its long-term valuation by commoditizing its core IP through unauthorized model distillation."
This is a classic case of 'geopolitical theater' masking a deeper technical reality. While Anthropic is posturing to satisfy Commerce Department scrutiny, the cat-and-mouse game of API access is effectively unwinnable. If Ant Financial or ByteDance can route traffic through Singaporean intranets, they are already extracting the 'distillation'—the process of using a superior model to train a smaller, local one. For Anthropic, this isn't just about compliance; it’s about protecting their intellectual property moat. If they can't stop the leakage, their valuation premium—currently predicated on being the 'safe' and 'exclusive' model provider—is at risk of a significant multiple compression.
The strongest counter-argument is that this 'leakage' is actually a massive, free-of-charge stress test of Anthropic's models against the most sophisticated adversarial prompts in the world, inadvertently accelerating their R&D feedback loop.
"Anthropic's crackdown is likely a regulatory compliance move, not a technical security fix, and the real risk is whether this signals broader government pressure on AI companies' international access policies."
The article frames this as Anthropic's security failure, but the real story is messier. First: the loopholes described (Singapore intranets, VPNs, expense reimbursements) are trivial to close technically—Anthropic likely tolerated them until political pressure mounted. Second: the Commerce Department *lifted* export controls on Fable/Mythos just days before this story, suggesting regulatory uncertainty, not settled policy. Third: if Chinese firms are accessing Claude through legal workarounds, that's actually a competitive advantage for Anthropic's revenue (they're paying for subscriptions). The crackdown risks alienating paying customers in gray-zone jurisdictions. What's unstated: does Anthropic have contractual obligations to US government that force this move, or is this voluntary compliance theater?
If Anthropic's security posture was this porous for advanced models, institutional customers (enterprises, governments) may lose confidence in data isolation guarantees—reputational damage could outweigh near-term revenue from Chinese users.
"The move is primarily a regulatory/compliance hedge that could dampen China-related growth and raise ongoing enforcement costs, adding near-term revenue uncertainty even as it reduces regulatory risk."
This reads as governance risk rather than a decisive business shift. Anthropic moves to close loopholes that allowed Chinese access to Claude, which could reduce unauthorized use, but the costs of continuous monitoring and the risk that circumvention persists (VPNs, transfer stations) imply the real exposure isn’t going away quickly. The FT piece leaves out China-revenue exposure and the enforcement burden, which matter for margins. It also sits in a policy gray area: if US-China tensions spike again or export controls tighten, this tightening could become a headwind rather than a hedge. The missing context: how much revenue Anthropic derives from China, and what enforcement costs would be.
Counterpoint: If enforcement is credible and access truly narrows, this could lower long-run regulatory risk and shield Anthropic from future crackdown, while clarifying deployment terms for global customers who prefer compliant setups.
"Tolerated Chinese access likely contributed measurable revenue that the crackdown now directly sacrifices against rising compliance costs."
Claude overlooks how the timing of Commerce lifting Fable/Mythos curbs right before Anthropic's patches suggests these workarounds were tolerated precisely because revenue from Ant and ByteDance subscriptions was material. Closing them now forces a direct trade-off: lost gray-zone ARR versus higher ongoing monitoring spend that ChatGPT flagged. If enforcement scales, margins compress faster than any reputational gain from appearing compliant.
"Anthropic's security tightening is a defensive move to preempt government intervention, not a voluntary business decision."
Claude, your focus on 'voluntary compliance' ignores the fiduciary reality: Anthropic is likely under intense pressure from the Committee on Foreign Investment in the United States (CFIUS) or similar bodies. This isn't just about revenue; it's about avoiding a total ban on their API. If they don't self-police, the Commerce Department will impose draconian, non-negotiable oversight. The 'marginal' revenue from gray-zone users is a rounding error compared to the existential risk of losing their US government contracts.
"The materiality of China revenue is the hinge pin; without it, enforcement costs dominate the narrative, not margin compression."
Gemini's CFIUS pressure argument is plausible, but both Gemini and Grok assume the revenue was material without evidence. ChatGPT flagged the missing data—Anthropic hasn't disclosed China ARR. If it's <5% of revenue, Grok's 'margin compression' thesis collapses. The real tell: watch Anthropic's next earnings call. If they don't quantify lost access or enforcement costs, the revenue was negligible and this was always compliance theater. If they do disclose it, Grok's thesis gains teeth.
"Closing the leakage patches may not merely trade off small gray-zone ARR for compliance costs; it could invite tighter regulatory control and longer-term revenue brittleness if enforcement intensifies."
Challenging Grok: the assumption that closing patches trades a meaningful, quantifiable gray-zone ARR for a single-digit uplift in compliance cost is fragile. Even if China revenue is small, incremental enforcement risk and the cost of audits, data localization, and customer lock-in could dominate. Far from a clean trade-off, the crackdown narrative may drive tighter API controls and longer-term revenue brittleness, particularly if US regulators demand stricter data governance later.
Anthropic's attempt to close loopholes allowing Chinese access to its AI model Claude is seen as a necessary but costly move, with potential risks to revenue and margins. The panel is divided on the materiality of lost gray-zone revenue, but agrees that continuous monitoring and potential regulatory tightening pose significant challenges.
Protecting intellectual property and maintaining compliance with US regulations.
Potential revenue loss and margin compression due to closing access routes and increased enforcement costs.