Update (Oct. 28, 2026): An earlier version of this post said no kill-switch legislation had been introduced. That was wrong. The AI Kill Switch Act (H.R. 9917) was introduced in the House on July 23, 2026, eleven days before this post was first published. We have corrected the post and added a section on the bill. We also corrected several details about the shutdown to match the sources.
In June 2026, the Commerce Department ordered Anthropic to suspend all access to its Claude Fable 5 and Mythos 5 models by any foreign national, three days after the models launched. Because Anthropic could not verify users' nationality in real time, complying meant taking both models offline for everyone, everywhere, according to the Center for Data Innovation. Fable 5 stayed offline until July 1.
The Center for Data Innovation argues that Congress needs clear, transparent rules for when and how the government can force an AI model offline.
What happened
Anthropic launched Fable 5 on June 9, Forbes reported. On the evening of June 12, Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei a letter directing the company to suspend all access to Fable 5 and Mythos 5 by any foreign national, anywhere in the world, including foreign nationals working inside the United States. "The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance," Anthropic said. Its other Claude models stayed available.
The administration issued the order after learning of a jailbreak that would allow Fable 5 to identify software vulnerabilities and produce exploit code, the Center for Data Innovation wrote. British lawmakers noted that hospitals, companies and researchers were using Fable 5 when it went dark, Fortune reported.
Amazon CEO Andy Jassy reportedly told Treasury Secretary Scott Bessent and other officials that Amazon researchers had used Fable 5 to obtain information that could be used in cyberattacks, according to Semafor reporting cited by Forbes. Fortune also reported that Jassy raised concerns with senior administration officials. White House adviser David Sacks wrote on X: "The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused." Anthropic disputed the severity of the jailbreak and said "the level of capability demonstrated is available from other publicly deployed models." According to the Center for Data Innovation, the government's evidence was presented verbally, and Anthropic said it never received disclosure of a jailbreak that produced a harmful result.
Commerce lifted the controls on June 30, CNBC reported. Fable 5 returned globally on July 1, and Mythos 5 was restored to roughly 100 vetted U.S. institutions. Our coverage of the July 1 return describes the new safeguards that came with it.
The policy problem
The Center for Data Innovation, a technology policy think tank, published an analysis by Michelle Lopes Maldonado on August 3 arguing that the episode exposed the lack of clear rules for when the U.S. government can restrict a deployed AI model. Its main points:
- No clear rules of engagement. The government's evidence was presented verbally, and public accounts from officials and Anthropic conflict. Outside the negotiating room, the rules were "ambiguous at best, and unknown at worst."
- No published standard. The reinstatement was a negotiated settlement conducted through private letters. It produced no published standard for what triggered the action or what satisfied the government's concern.
- Cost to users. The shutdown came suddenly and without warning. The analysis expects careful buyers to price in the risk by building redundancy and splitting workloads across vendors.
- Strategic cost. For governments and businesses abroad, the lesson was that American AI is "a revocable dependency." The analysis calls that a gift to Chinese open-weight developers, whose models no U.S. authority can recall.
The day after the analysis ran, Bloomberg reported that China was drawing up retaliation options over the U.S. restrictions on Mythos.
The broader precedent
Forbes called the order one of the most aggressive uses of export-control powers against a commercially deployed AI model. Fortune reported that it was the first time the U.S. government had used export controls to halt access to a commercial AI model already widely used by the public.
One day after Fable 5 launched, Amodei published a policy essay calling on the U.S. government to hold legal authority to block or reverse frontier AI models that fail independent safety testing, Forbes reported. Two days later, the government used its export-control authority against his own company.
The shutdown also came at a sensitive moment for Anthropic. Forbes reported that the company had filed a confidential IPO prospectus with the SEC earlier in June, disclosing a revenue run rate of $47 billion and a valuation of $965 billion.
What Congress could do
The Center for Data Innovation says the administration can act first by adopting clear protocols. For hosted U.S. models, those would include written notice, an opportunity to respond and remediate, defined timelines and an appeals path. It argues that, absent a credible risk of catastrophic harm, the government should use proportionate remedies, such as a coordinated vulnerability disclosure process, before imposing a global shutdown.
If Congress acts, the analysis says it should codify the June 2, 2026, executive order "Promoting Advanced Artificial Intelligence Innovation and Security," protect confidential disclosures of severe AI vulnerabilities, and create a standardized CVE-style system for rating and disclosing AI jailbreaks. It also notes that a switch-off authority applies fully only to hosted, closed-weight models, because open-weight models cannot be recalled once released.
The AI Kill Switch Act
One bill was already pending when the analysis ran. Rep. Ted Lieu (D-Calif.) and Rep. Nathaniel Moran (R-Texas) introduced the AI Kill Switch Act, H.R. 9917, on July 23, 2026. Reuters reported the bill the same day. Based on the introduced text, the bill would:
- Require covered developers to maintain the technical ability to stop inference, terminate user access, suspend specific accounts or users, and shut down a covered AI system. Covered developers are those that offer such a system to third parties through an API or hosted service and earn at least $500 million a year in gross revenue from it. A covered system is one trained with compute that would cost more than $100 million at U.S. cloud prices.
- Require covered developers to report a "covered incident" to the Department of Homeland Security within 15 days.
- Let the Homeland Security Secretary, acting through the CISA director and in consultation with the Commerce Secretary and the Director of National Intelligence, order proportionate action, up to a shutdown, after determining that a covered incident has occurred. The developer would have to notify affected operators and users where practicable.
- Define a covered incident as sabotage of a lawful shutdown instruction, unintended conduct that causes at least 10 deaths or $100 million in damages, a system concealing its capabilities or actions from monitoring or shutdown mechanisms, or a loss-of-control scenario, in each case outside red-teaming or structured testing.
- Give a developer 48 hours to petition for reconsideration, without pausing the order, and allow review in the U.S. Court of Appeals for the D.C. Circuit.
- Allow civil penalties of up to $20 million per day for violating an emergency order.
The bill was referred to the House Homeland Security Committee and, on July 24, to its Subcommittee on Cybersecurity and Infrastructure Protection. Congress.gov lists that referral as the latest action. The bill gives DHS, not Commerce, the shutdown power, and it does not change the export-control authority Commerce used against Fable 5.
The debate has also moved to the states. In October, Anthropic told California that when a review finds a model poses a serious risk, the government should be able to ask a court to block it, as we reported in our coverage of Anthropic's letter to Governor Newsom.
Sources
- Center for Data Innovation: Congress Can Bring Clarity to AI Shutdown Authority (Aug. 3, 2026)
- Congress.gov: H.R. 9917, AI Kill Switch Act, all actions
- GovInfo: H.R. 9917 text as introduced in the House
- Reuters: White House monitors OpenAI's "rogue" AI incident, lawmakers propose "kill switch" (July 23, 2026)
- Forbes: Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order
- CNBC: Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5
- Fortune: How a warning from Amazon led the White House to shut down Anthropic's Mythos model
