Washington has restricted access to Anthropic's most powerful AI model for any foreign user, then opened a classified channel with India on how to share it safely. Power grids are the test case, and other US partners are watching.
The United States has, for the first time, used an export-control directive to restrict access not to the chips that run a frontier AI model, nor to its weights, but to the model's actual functionality for any foreign user anywhere in the world. It has then opened a quiet, classified-track channel with India to decide who else gets to use that AI, and under what conditions, with the early test case being whether the technology is safe to plug into power grids.
That is the operational shape of what US Under Secretary of Economic Affairs Jacob Helberg, a senior State Department economic-security official as identified by Indian wire reporting, described in a recent interview with ANI as "very sensitive national security discussions that are not quite right for public consumption." Helberg said Washington wants "a gradual, measured approach to how we release Anthropic's models in a way that is safe, both for ourselves but also for our Indian counterparts as well as all our trusted partners for our critical infrastructure, for our power grid."
The choice of venue matters. The conversation is being framed as critical-infrastructure diplomacy, not as product release or commercial deal-making. Fable, an Anthropic frontier model, has become the diplomatic test case for a much larger question: when the most powerful AI systems start running the systems that keep electricity, telecoms, and finance running, who decides who gets access, on what timeline, and with what reliability guarantees?
The trigger was a directive Anthropic itself disclosed. The company announced that "the US government, citing national security authorities, issued an export-control directive suspending all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, including foreign-national Anthropic employees." The directive forced Anthropic to disable the two models for every customer, US-based or otherwise. The story was first reported in Indian outlets including the Times of India and The Hindu Business Line.
Two things make this novel. Existing US export controls on AI, including the chip rules that have shaped Nvidia's China business, regulate hardware or model weights that physically cross borders. This directive does not. It restricts access to a model's outputs by foreign nationals who may be sitting in California, London, or Bengaluru, and it does so because of what the model can do, not where the server sits. Independent legal analysis and Cloud Security Alliance research both describe this as the first time an export-control regime has reached the intelligence layer of a frontier model itself.
That expansion is what makes the India channel consequential. Indian officials have previously raised reliability-of-access concerns with Western AI providers; in the latest round, the Indian side asked for assurances against "abrupt cutoffs," and was told trusted-partner access would not be an issue going forward, according to the wire reporting. The catch is opacity. Helberg's "not quite right for public consumption" framing means the substance of the negotiations, including any commitments on pricing, deployment timeline, or model specifications, is being kept out of public view.
The bilateral track also sets a template. If frontier AI is functionally treated as critical-infrastructure software, deployment decisions move out of normal product and export-control lanes and into a small set of bilateral national-security channels. That changes how other US partners, including Japan, the EU, and Gulf states, will encounter the same frontier models, and it puts India's own AI sovereignty and reliability-of-access asks at the center of an emerging US framework rather than at its margins.
What remains undecided is what comes out of the channel. Helberg offered no rollout date, no list of which models beyond Fable and Mythos fall under the new regime, and no public criterion for when an AI system crosses into "critical" status. The classified framing keeps those answers out of view for now, which is itself part of the story: the most consequential frontier-AI coordination between Washington and a strategic partner is being conducted in a form designed to resist public scrutiny.