President Donald Trump ordered US agencies to work with artificial intelligence companies on protecting against AI-enabled cyber attacks under a long-awaited directive that he had postponed last month citing his concern that it risked stifling tech innovation.

The executive order signed Tuesday would grant the US government access on a voluntary basis to cutting-edge AI models to test their capabilities up to 30 days before their release to the public, with measures to ensure confidentiality. It also calls for making it easier for federal, state and local agencies as well as operators of critical infrastructure to access cybersecurity tools embedded in so-called frontier AI models.

Under the directive posted on the White House website, the government would also establish a clearinghouse for industry and agencies, including the Treasury Department and Pentagon, to coordinate and deconflict software vulnerabilities. Trump also assigned the National Security Agency and other agencies to work with AI developers on standards to determine whether an AI model was advanced enough to qualify under the order.

The highly anticipated policy had been a source of debate within the administration, as officials weighed how much of a role the US government should play in testing new AI models before they are released to the public. Trump was scheduled to sign the order in late May but scrapped plans for a signing ceremony at the last minute, telling reporters he “didn’t like certain aspects” of the directive and worried it would undercut US competition against China in the AI race.

Trump had faced pressure from the tech industry over elements of the order ahead of its release. His former AI and crypto czar, David Sacks, had raised objections personally to the president, according to several media reports.

Tuesday’s order emphasizes voluntary participation by developers despite earlier deliberations about mandating government approval of new AI systems. It shortens the government review period for models to a maximum of 30 days, down from 90 days under an earlier draft.

The AI security push comes after Anthropic PBC revealed that its new Mythos model was extraordinarily adept at finding network vulnerabilities and could pose a major cybersecurity risk. The company had limited Mythos access to a handful of large technology and Wall Street companies, amid broader global alarm about the new threats it could pose, but on Tuesday it announced that the model would be made available to 150 additional organizations.

Read more: Anthropic Offers Mythos Model Access to 150 Additional Groups

The Mythos breakthrough prompted the administration to accelerate existing efforts to craft AI policy that would address a range of security questions, an undertaking led by White House Chief of Staff Susie Wiles. Wiles and other administration officials met on April 17 with Anthropic Chief Executive Officer Dario Amodei, where topics discussed included Mythos.

The Trump administration had sought to make Mythos more widely available to federal agencies to test their networks for security flaws. US government officials and cybersecurity experts have expressed concern that future models will continue to improve and ultimately be used to threaten national security.

The order makes clear the Pentagon must take action to protect its information systems by collaborating with the private sector, without giving further detail on which companies it would have to collaborate with. The Defense Department has been embroiled in a bitter dispute with Anthropic over usage guidelines the company insisted be included in its government contracts. Defense Secretary Pete Hegseth moved to remove the AI company from all military workflows, but other parts of the government, including the NSA, have continued working with Anthropic.

Under the directive, AI developers could submit so-called frontier models to the government on a voluntary basis, to promote cybersecurity. The measure provides no specifics on safety testing.

Currently, the US Commerce Department already runs a voluntary program to evaluate AI systems before their release. Though Commerce is listed as one of the agencies involved in benchmarking, the order says that the director of the NSA will make decisions about which models may be included in pre-release government evaluations, in consultation with others.

As part of the existing Commerce Department program, Alphabet Inc.’s Google, Microsoft Corp. and xAI have agreed to give the US government access to their models. OpenAI and Anthropic were already part of the initiative, led by the Center for AI Standards and Innovation.

Read More: NSA Testing Anthropic’s Mythos to Find Flaws in Microsoft Tech

In the two months since Anthropic announced Mythos and said it was too powerful for a public release, the model has found 6,202 high and critical-severity vulnerabilities in open source software, according to the company. Anthropic has long said that other labs would soon catch up. Tests by the UK government’s AI Security Institute show that ChatGPT is one of the strongest models across cybersecurity-focused tasks such as reverse engineering, web exploitation and cryptography.

— By Courtney Subramanian and Maggie Eastland (Bloomberg News)