WASHINGTON -- The Trump administration and Anthropic have reached an agreement to restore access to the company's most recent general-access artificial-intelligence model, resolving a fight that showed how the White House is intervening to address security concerns in the fast-growing industry.
Under a deal announced Tuesday evening, Anthropic would address the workarounds that researchers at Amazon used to evade the safeguards for Fable, a public version of Anthropic's powerful Mythos model that is capable of carrying out cyberattacks, Commerce Secretary Howard Lutnick said on X. The guardrails are critical for the company to release the model publicly.
"Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5 to ensure alignment across the U.S. government and strengthen America's leadership in AI," Lutnick said.
Anthropic said it is working with other tech giants including Amazon, Microsoft and Google to develop a consensus framework for gauging the severity of so-called jailbreaks that bypass safety measures and how AI developers should respond. The effort is a sign that the recent Anthropic model shutdown will shape future AI development and government oversight. Anthropic said that it was "probably impossible" to make any model jailbreak-proof.
In recent talks led by Anthropic's Chief Compute Officer Tom Brown, the company worked with the Commerce Department and other agencies to convince them that the model is safe for the public.
Following the agreement, Anthropic said it would begin restoring access Wednesday after a 2 1/2 -week shutdown that was unprecedented for a leading U.S. model developer. "We're grateful to our users for their patience, and to everyone who worked with us on redeploying," the company said.
Anthropic implemented a new safeguard to address the behavior Amazon researchers flagged after working closely with the government, according to the company. The technique described by Amazon to evade guardrails fails about 99% of the time with the new safeguard, according to Anthropic. For the other cases, the outputs aren't considered risky because they contain publicly available information or couldn't help a bad actor, Anthropic said in a blog post.
The company previously had said it would redirect prompts submitted to Fable that were deemed risky to a less advanced model.
The government's goal was to address the vulnerability and implement a process so similar issues can be avoided in the future and communication is clearer, administration officials have said. U.S. officials recently approved some access being restored for a related model called Mythos 5 after a two-week shutdown.
Concerns about the safety of the models prompted the administration to ban all foreign use of the tools in mid-June. The decision included foreign-born individuals working in the U.S. and some Anthropic researchers, prompting the company to shut off access for all users. The Commerce Department is expected to lift the restrictions, people familiar with the matter said.
The model shutdown marked a seminal moment for the AI industry, the first major instance of the government forcing a leading AI company to take down a model and potentially benefit that firm's competitors. ChatGPT maker OpenAI has products with similar capabilities that are being released on a limited basis at the government's instruction. Google and Elon Musk's AI company are also competing in the space.
After implementing a laissez-faire approach for more than a year, the White House has pivoted, wanting cybersecurity and national-security experts more involved in evaluating AI models and threats they may pose.
President Trump recently signed an executive order laying out the new approach and asked companies to give the federal government access to their models 30 days before they are released.
"The government and private sector have worked together in a way we have never seen before and this foundation of America first is unprecedented, " White House chief of staff Susie Wiles said on X Tuesday. "Our shared priority remains: get the best tech deployed as quickly and safely as possible."
Trump recently praised Anthropic Chief Executive Dario Amodei's response to the administration's concerns that led to the shutdown, telling Axios, "He responded very responsibly." The pair took part in a Group of Seven AI meeting in France while negotiations were ongoing. Earlier this year, Trump called Anthropic a "radical-left, woke company" when the Pentagon designated the company a security risk over disagreements about guardrails for the use of AI in the military.
Potentially resolving the latest spat is a relief to many AI analysts who said Anthropic and the administration have a self-defeating rivalry. The Pentagon's designation of the company as a supply-chain risk prompted multiple lawsuits. The two sides have also clashed over exports of AI chips to other countries and the company's ties to nonprofits that are big liberal donors.
AI analysts say the administration's recent moves have thrown the model approval process into chaos. A government unit already does testing that security officials will now oversee and has long relationships with tech companies. The Anthropic agreement is also expected to involve the government testing unit, the Center for AI Standards and Innovation.
CAISI tested the new safeguard in addition to the prior guardrails, according to Anthropic.
Industry executives and administration officials have said some of the uncertainty will be resolved after the executive order is implemented.
The agreement could also relieve pressure on Amazon, which is a big Anthropic investor and partner on data centers that run Anthropic's models but flagged vulnerabilities in Fable to the government for security reasons. The e-commerce titan has invested $13 billion in Anthropic and could put in $20 billion more. Anthropic was recently valued at $965 billion and is racing toward an initial public offering alongside OpenAI.
Write to Amrith Ramkumar at amrith.ramkumar@wsj.com and Robert McMillan at robert.mcmillan@wsj.com
(END) Dow Jones Newswires
June 30, 2026 23:46 ET (03:46 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments