Anthropic has stoked fears about artificial intelligence more aggressively than any other major AI company. It has repeatedly urged governments to create “authority with teeth” to block unsafe AI models. On June 21, the government did exactly that, using export-control authority to force Anthropic to block access to Fable 5 and Mythos 5.
Some see the government’s move as poetic justice. But Anthropic may have lost a skirmish only to win the war. The government’s Mythos action was rushed, opaque, and legally dubious, not least because it relied on narrow export control authorities to ban the use of an American model by Americans at home. With industry already pining for a more predictable legal process, Anthropic could get exactly what it has wanted all along: a regime whereby the government must bless new AI models before the public can use them.
Recommended Stories
That would stifle competition, raise prices, and drive the global market toward free and increasingly capable Chinese open-source models. There is a better way.
The government’s Mythos action raised fears that a new White House process for voluntary pre-release review would prove mandatory in effect. Sure enough, OpenAI CEO Sam Altman recently informed staff that the government will be approving customer access to its newest model, GPT 5.6, on a case-by-case basis. The government demanded to review GPT 5.6 before its release, and seeing what happened to Anthropic, OpenAI felt it could not say no.
Everyone agrees that the White House’s current approach is unsustainable. Export controls are a poor fit to control software that millions access remotely, overbroad when applied to foreign nationals inside U.S. firms and allies abroad, and so arbitrary that they will inevitably undermine global trust in American AI. As the administration seeks a better process, Anthropic has one ready: an FDA-style approval regime for frontier AI models, which the company argues will be transparent, safe, and narrow.
It’s a familiar regulatory ratchet. First, the government seizes control in an emergency. Then industry incumbents ask it to formalize that control. Finally, their compliance departments, lawyers, and lobbyists turn the resulting rules into a moat. The largest firms can survive months of uncertainty and armies of evaluators. Startups, university labs, and, crucially, open-source model developers cannot. “Cybersecurity” becomes an excuse for slowing everyone to the pace at which the government can patch its systems.
We’ve seen this story before. A century ago, AT&T “settled” with the Department of Justice to cement their monopoly. Similarly, the tobacco industry embraced the 1998 Master Settlement Agreement to create a legalized cartel, allowing it to reduce output and raise prices without violating antitrust laws.
This lose-to-win strategy will hurt most in AI. Delaying the release of frontier models delays their economy-wide benefits. It slows progress toward better medical diagnosis, faster drug discovery, more secure hospitals, cheaper legal services, stronger fraud detection, and productivity growth across the economy. Yet any pre-release approval process is likely to undervalue these benefits because no one can measure them in advance, while the cyber threats look concrete. The result is a risk-averse process that will forfeit the broad benefits of AI in the name of cybersecurity.
Except that process won’t even make us more secure. Mythos-level cyber capabilities do raise the stakes. But government, infrastructure operators, and private firms have managed cyber risk for decades, and that experience teaches one lesson: Security is iterative. You cannot certify software “secure” in the lab. Defenders test and harden software in the wild, which is why Windows and macOS ship patches more often than any of us gets a haircut. A pre-release approval queue for frontier models would bottleneck that iterative cybersecurity process, blocking American defenders’ access to our best tools while attackers use increasingly capable free Chinese models.
The right framework lets competition harden systems as models grow more powerful, and it divides the work. Labs track and prevent harmful uses. American-led open-source AI gives users alternatives to Chinese models. Agencies help critical infrastructure patch faster. Law enforcement pursues criminals who weaponize AI. Congress builds federal capacity for rapid threat assessment, classified information sharing, vulnerability clearinghouses, and coordinated fixes.
That framework makes defenders quicker to adapt and systems harder to attack. It does not slow AI innovation to the speed of government IT. It does not require Americans to ask permission to use new models.
US RELEASES ANTHROPIC’S MYTHOS MODEL TO COMPANIES AND AGENCIES AFTER BLOCK
Nor does it subject today’s AI companies to a regime that, although painful, ultimately favors the current market leaders.
Anthropic may look like today’s loser. We should worry that this episode could make it tomorrow’s government-picked winner.
Neil Chilson is the head of AI policy at the Abundance Institute and former chief technologist at the Federal Trade Commission.
