Whatever went on in closed-door negotiations between Anthropic and the US government, it worked. Foreign access restrictions on the company’s Fable 5 and Mythos 5 artificial intelligence (AI) models have been dropped after Anthropic resolved the Trump administration’s safety concerns.
The startup’s long-time focus on investing in safety had made the initial ban sting. What is more unsettling now is that the agreements it made with the government to secure the reversal are largely a mystery.
The ban was triggered when Amazon told the US administration about concerns that certain safeguards on Fable could be circumvented. Now Anthropic has built a way to block that workaround and the fix was supported by the Commerce Department’s Center for AI Standards and Innovation (CAISI).
It did this via so-called classifiers, or smaller AI systems that detect when a request is about prohibited cybersecurity tasks. Anthropic has not disclosed more details about the changes it made beyond the broad mechanics of the fix.
But what about future AI systems that could also be dangerous if weaponized? It seems Anthropic and the government hashed out a rough plan for that too. US Commerce Secretary Howard Lutnick said the company had agreed to collaborate on “protocols and standards” for upcoming models.
Anthropic is also working with other tech companies, including Amazon and Microsoft, on developing those protocols. They should be aired while they are still being written and not unveiled once they are settled, when public scrutiny can no longer help shape them.
A long-necessary set of security standards for AI models will be good for the industry, and that is the silver lining to the messy nightmare that the export ban caused for Anthropic.
The company had already been proactively detecting risks and giving the CAISI access to its models before release. Now, according to a person close to the company, it will send its staff to CAISI’s offices and offer the agency computing power and its own safety testers for deeper analysis. Whatever previous collaboration it had offered regulators, it will start doing more.
As for the US government’s opacity on the agreement, plenty of export controls and national-security deals are kept under wraps. What is unusual here is the improvised, case-by-case nature of a public ban on a novel product from a company that just happens to have taken an adversarial stance against the government—a ban communicated by private letter that lasted for 18 days before being reversed after closed-door negotiations.
The framework to come should at the very least be out in the open, otherwise the public has no way of knowing whether AI is being vetted on legitimate national security grounds or because of favouritism. Even Anthropic archrival OpenAI seems to agree. When it released its latest model GPT-5.6 to a short list of government-approved partners, its CEO Sam Altman said he disliked the idea of regulators “picking the customers.”
Though radical, it would make more sense to outsource security testing of leading models to a government-run agency in the UK, known as the AI Security Institute, which is far better funded and staffed with more engineering talent than the CAISI, its US counterpart. A British agency has a further advantage: It avoids the risk of companies currying political favour with those in charge of oversight.
Crucially, the standards should be public, so that enterprise customers of Anthropic and OpenAI that are already becoming more concerned about costs do not have to worry about a kill-switch risk too. The ban on Anthropic deeply unnerved companies in Europe, prompting calls to invest in homegrown AI so they do not rely on American models that can be ripped out from under them without warning.
It is possible Anthropic’s hidden concessions were ultimately a face-saving veneer for a retreat by the White House itself, sparked by concerns raised by the tech industry over the last two weeks. More than 100 cybersecurity experts had signed an open letter arguing that the export ban was excessive, and multiple tech executives and investors had publicly warned that it would give further time for Chinese AI developers to gain a greater share of the market.
Whatever the reason, AI governance should not be an improv act, but should instead be laid out in clear and public standards. With the US reversal, Anthropic can celebrate the release of its most sophisticated product. But the rest of the industry should use this moment to call for greater clarity on what could trigger yet another blockade—or the uncertainty already threatening to engulf the market for AI will only get worse. ©Bloomberg
The author is a Bloomberg Opinion columnist covering technology.