Cloud Security Act Targets AI Loopholes, But Risks Handcuffing US Defenders

The government wants to know when a foreign actor uses cloud AI to build advanced models, but the same impulse just yanked Anthropic's best product offline and made OpenAI gatekeep its next release—turning defensive AI into a political bargaining chip.

On June 26, Representatives Josh Gottheimer (D-NJ) and John Moolenaar (R-MI) introduced the bipartisan Cloud Security Act. The bill would require AI companies to flag suspected "misuse" of their platforms to the federal government. Backers say it closes a loophole in existing export rules meant to keep advanced AI chips out of rival countries like China. The rationale is sound on paper. But the real story is what's already happening without any new law.

The Commerce Department Already Polices Models, Not Just Chips

We're not talking hypotheticals. In recent weeks, the Commerce Department worked with OpenAI on a deal that sharply restricts which customers get access to the upcoming ChatGPT 5.6 model. Separately, the department pushed Anthropic to roll back its release of Fable 5, a version of its Mythos model, citing security concerns. Anthropic complied and took the model offline. That's effectively a government veto on a company's most advanced product—no public oversight, no clear standard.

This isn't just about Anthropic or OpenAI. Representative Sam Liccardo (D-CA), one of four lawmakers who sent a June 18 letter to Commerce about the Fable 5 takedown, put it bluntly: "If we're going to take the defensive capability out of the hands of U.S. government, of local government, and of U.S. industry, and we're going to tie our hands... we better have a really good and clear justification for that because the downside risk is enormous." His office says the agency missed the June 26 deadline to respond to their questions.

The Real Risk: Handcuffing Domestic Defenders While Adversaries Race Ahead

Liccardo's concern isn't about one model or one company. If adversaries are building increasingly powerful AI, American institutions—from federal agencies to local governments to private industry—may need access to the most capable U.S. models to defend themselves. A blunt export or access restriction could leave those defenders without the very tools they need to counter attacks. Worse, there's no transparent standard for when or how these limits are applied. "If we have an export licensing regime without any review or substantial oversight until months or even years later," Liccardo warns, "that oversight becomes worthless and meaningless in an industry where progress is measured by days and weeks."

The Cloud Security Act might close one loophole, but it doesn't address the core tension: the same government that wants to track AI misuse is already picking winners and blocking products without a clear playbook. Liccardo's clock is ticking.

Source: The government wants to rein in powerful AI, but there are downsides
Domain: fastcompany.com