Access Anthropic Security
AI Access as National Policy: The Anthropic Ban, Chinese Rivals, and a Novel DOJ Legal Theory
The US government, exercising export control authority, blocked international access to Anthropic's most advanced AI models — Fable 5 and Mythos 5 — while negotiations between Anthropic and the White House over carve-outs or managed access frameworks collapsed without a deal. The fallout has reached the highest levels of allied diplomacy: UK Prime Minister Starmer and Canadian Prime Minister Carney are heading to the G7 summit specifically to confront the Trump administration over restrictions that treat close allies the same as adversary state actors in terms of AI access.
The administration's counterargument — that ally exemptions create pressure for further exemptions and allow models to migrate through intermediaries to restricted destinations — is not without logic. But blanket restrictions without a trusted-partner framework are generating real economic and research disruption among genuinely aligned countries. Anthropic's own workaround, an identity verification policy taking effect July 8th that restores access to verified US citizens, sets a new precedent: access to AI capabilities gated by citizenship documentation rather than the terms-of-service agreements that have historically governed software access. The privacy implications of collecting identity documents at that scale remain largely unexplored.
The DOJ filed a separate, striking argument this week: that xAI's data center — Elon Musk's AI company — is vital to national security and should therefore be immune from an NAACP environmental pollution lawsuit filed by affected communities. The legal theory, if accepted by courts, would effectively allow companies with government contracts to weaponize national security designation as litigation immunity against environmental accountability claims — a precedent legal observers say appears to be without clear prior precedent.
Chinese AI competitors are not pausing while Washington works out its export control framework. DeepSeek raised $7.4 billion in its first funding round at a valuation above $50 billion, demonstrating that the Chinese AI ecosystem retains substantial domestic and international capital access. Zhipu, another Chinese AI developer, saw its stock surge 48% in Hong Kong after JPMorgan raised its price target. The premise of US export restrictions is partly to preserve American competitive advantage; DeepSeek's fundraising suggests the competitive gap may be narrower than policymakers assumed.
A Cornell University study added a different dimension to AI security concerns: researchers found that malicious content planted on Reddit can manipulate AI agents that use web search as a tool. The agent retrieves adversarially crafted content, treats it as credible, and produces outputs or takes actions based on poisoned information — prompt injection at scale. The finding suggests that agentic AI systems carry vulnerability surfaces that have not yet been fully mapped, even as those systems are being deployed in increasingly consequential applications.