The US Department of Commerce has lifted export restrictions on Anthropic’s two most advanced AI models, Claude Fable 5 and Mythos 5, just weeks after ordering their suspension on national security grounds. As BBC News reports, the models were abruptly pulled from global access on June 12 amid concerns they could be used to exploit vulnerabilities in computer systems. Commerce Secretary Howard Lutnick confirmed in a letter to Anthropic that the company “has addressed the risks” by agreeing to proactively detect and address security vulnerabilities, collaborate on future model releases, and alert the government of any malicious activity detected. The restrictions applied to Fable 5, a consumer-facing reasoning model, and Mythos 5, a business and cybersecurity-focused model capable of identifying and exploiting code vulnerabilities. The Department reserves the right to reimpose restrictions if necessary.
The received wisdom
The tech industry’s response to episodes like this tends toward indignation about government overreach. The standard objection is that security agencies lack the technical expertise to assess AI models meaningfully, that export controls on software are structurally different from controls on physical weapons, and that restricting access to a commercial AI product deployed to hundreds of millions of people — as Anthropic itself noted — is operationally incoherent as a national security measure. On this account, the government’s twenty-day reversal proves the original order was ill-considered: the same concerns that justified the suspension were apparently resolved by a voluntary compliance agreement that Anthropic would presumably have offered at any point. The episode is presented as a cautionary tale about security bureaucracies that reach for the regulatory lever without understanding the technology they are attempting to control.
A different read
The indignation is understandable, and the procedural complaint has some merit. But the casual dismissal of national security review of frontier AI models as mere bureaucratic overreach is, on examination, considerably less persuasive than it appears.
Start with the specific capabilities at issue. Mythos 5 is explicitly described as a model designed for cybersecurity experts, capable of identifying vulnerabilities in computer code and exploiting them. That is not a description of a chatbot that helps users draft emails. It is a description of a system with meaningful offensive cyber capabilities that can be directed by anyone with API access, anywhere in the world, at any target. The concern that a method of “jailbreaking” Fable 5 had been identified — circumventing the safety restrictions that are supposed to prevent misuse — is precisely the kind of concern that any responsible government should take seriously, even if the initial response was imperfect.
Anthropic’s objection at the time of the suspension — that “the finding of a narrow potential jailbreak should not be cause for recalling a commercial model deployed to hundreds of millions of people” — is revealing in a way the company perhaps did not intend. The argument is essentially that the scale of a deployment creates its own immunity from security review: the product is so widely used that stopping it is too disruptive to contemplate. That argument, if accepted as a general principle, would effectively preclude any national security action against any widely deployed technology, regardless of the risk. It would also be familiar to anyone who has watched large financial institutions argue that they are too systemically important to be regulated stringently.
The resolution — a voluntary compliance agreement under which Anthropic commits to security monitoring, government collaboration, and incident reporting — is actually a more mature outcome than either a permanent ban or no action at all. It resembles, loosely, the framework that governs other dual-use technologies: the government retains oversight authority, the company retains commercial operations, and there is a defined channel for addressing security concerns as they arise. As BBC News reports, the Commerce Department explicitly reserves the right to reconsider if new concerns emerge.
The deeper problem is that this framework is being constructed improvisation by improvisation, without a legislative mandate, an established agency structure, or settled legal principles. The Lutnick letter resolves one episode; it does not create a durable architecture for governing AI capabilities with genuine dual-use potential. The next AI model with greater capabilities will face the same improvised review by the same unprepared institutions — unless Congress creates the framework that currently does not exist.
There is a conservative case for government restraint in technology regulation that takes innovation and economic dynamism seriously. But there is equally a conservative case for the proposition that the state has legitimate interests in controlling technologies with genuine weapons-grade potential, and that those interests should be institutionalised in law rather than handled through ad hoc Commerce Department letters. The current approach satisfies neither principle.
What to watch
Track whether any legislative proposal for an AI national security review framework advances in Congress following this episode — the Lutnick letter creates a precedent but no legal structure. Watch Anthropic’s compliance agreement disclosures: if the company is required to publish any information about the security monitoring regime it has agreed to, that will reveal how substantive the oversight actually is. The next frontier model release from any major lab — OpenAI, Google DeepMind, or xAI — will test whether the Commerce Department applies consistent scrutiny or whether Anthropic was singled out. The international dimension is the most significant long-run question: will US allies adopt compatible oversight frameworks, or will divergence create regulatory arbitrage that routes the most sensitive AI development to less scrutinised jurisdictions?
— J