WASHINGTON — For America’s AI industry, the message in President Donald Trump’s new memo on national security is unmistakable: Uncle Sam wants you — unless you’re Anthropic.
Of course there’s an exception to that exception: Anthropic’s cutting-edge Claude Mythos, which is in use at the NSA, and for which the memo allows a waiver, showing just how complicated the administration’s relationship with tech titans has become.
Overall, National Security Presidential Memorandum (NSPM) 11, issued Friday, aims to accelerate adoption of AI in national security. It seeks to streamline procurement, expand training, and enhance security in close collaboration with America’s leading tech firms, while revoking Biden-era guardrails it sees as too restrictive. But the memo also orders the “termination for default or for convenience” of contracts with any company that tries to restrict how the government uses its AI, as the administration alleges Anthropic did, with strictly limited waivers allowed for pressing needs, as reportedly is the case at NSA.
The memorandum is careful not to name specific companies. Nonetheless, AI and legal experts who spoke to Breaking Defense agree, as do many reputable observers, that the six-page policy edict is unmistakably shaped by the administration’s ongoing legal battle with Anthropic and its desire to strengthen its position for any future disputes with its AI contractors.
“Most of the attention will go towards the [memo’s] discussions about working with companies, which clearly reflects the ongoing dispute between the Trump administration and Anthropic,” former Pentagon policy official Mike Horowitz told Breaking Defense.
“There’s absolutely no question whatsoever, this comes from the Anthropic dispute,” agreed Jack Shanahan, who founded Project Maven and the Pentagon’s Joint AI Center. “There’s no other interpretation,” he said in an interview.
Anthropic’s Claude was the first commercial Large Language Model authorized for use on classified networks. But in February, the company questioned the reported use of Claude to plan combat operations against Venezuela and Iran. The administration retaliated by starting to switch to other companies’ AIs while cancelling all of Anthropic’s federal contracts on the grounds it was a “supply chain risk,” an action the company challenged in two parallel lawsuits.
Experts are divided on how the new memo impacts these ongoing cases. As a veteran of federal AI procurements, Shanahan thought the memo comes too late to affect a judge’s ruling — which must be based on what held true at the time a suit was filed — but could give the administration a “do-over.” If it loses the lawsuits, he told Breaking Defense, the administration could invoke the new policy directives in NSPM-11 as the basis to try cancelling Anthropic’s contracts again.
A lawyer specializing in federal procurement, however, thought that scenario would be a legal longshot.
“I don’t know their intent, but I never read this as a litigation hedge,” said Jessica Tillipman, associate dean at George Washington University. Instead, she argued, the memo seems to be looking at what went wrong with Anthropic and trying to shore up the Administration’s position for the next fight, whenever and with whomever that may be.
“It’s less about specific Anthropic [lawsuits] and more about an ongoing challenge within the government, which is dependency on single suppliers,” she told Breaking Defense.
If the administration wants this reform to stick, Tillipman warned, it’s not enough to issue this sweeping policy memorandum. It also has to implement it with individual companies, contract by contract.
“It’s clear they want to be able to get rid of a contractor easier, including subcontractors,” Tillipman said. “I’m waiting to see what that looks like in a clause.”
The specific wording of specific clauses is crucial, agreed the former CTO of the CIA, Stanford Hoover Institution fellow Nand Mulchandani.
“The Anthropic thing was a contract issue that got out of hand,” he told Breaking Defense. “What this [memo] is doing, basically, is guiding officials who are negotiating the contracts, but it depends on the contract you write.”
This uncertainty about contract language, moreover, is just one way in which the ultimate impact of this memorandum will depend on the devilish details of how it’s implemented. The other big unanswered question: What happens to the Pentagon’s most important policy document imposing safety guardrails on military AI?
Unanswered Questions On “Responsible AI”
While NSPM-11 urges both government and industry to move faster to implement AI, the memorandum also emphasizes legal accountability and technical testing.
“The United States can and will responsibly accelerate the use of AI across intelligence and warfighting domains in line with American values,” the memo says. “We will streamline the acquisition and deployment of these technologies while maintaining rigorous oversight … including testing, evaluation, validation, and verification.”
While Trump’s memo takes time to denounce the Biden administration’s AI policy memorandum, NSM-25, Shanahan told Breaking Defense, “I don’t see it as a dramatic break. It does emphasize the right things about testing, steerability, and controllability.”
Horowitz, who helped author those Biden-era policies, agreed. “If you look past the rhetoric and the Trump Administration flourishes, there’s a lot more continuity here than change,” he told Breaking Defense. “Even some of the changes are pretty common-sense. For example, updating the Pentagon directive on autonomous weapon systems was already announced, and it makes sense for policy to adapt as the technology changes.”
But how the safeguards actually get implemented will depend heavily on the details of that updated directive, DoDD 3000.09 [PDF], “Autonomy in Weapons Systems,” which was last updated under former President Joe Biden in 2023. Trump’s memo tasks Defense Secretary Pete Hegseth to submit a new version of 3000.09 within 90 days.
That’s a tight deadline, said Shanahan, a veteran of such Pentagon policy deliberations, but it’s entirely doable — if the rewrite is limited to modest changes. The current version of 3000.09 was written shortly after the release of ChaptGPT and before the current explosion of AI, he noted for example, so those aspects could definitely use an update.
A more sweeping overhaul, however, would be hard to do properly in three months, Shanahan warned. “It all depends on the extent of the changes that they’re asking for,” he told Breaking Defense. “You’re going to see a lot of fingers crossed out there in the tech companies, waiting to see, is this a reasonable revision to 3000.09? If it is, they’re going to breathe a sigh of relief.”
“Speed is great,” Shanahan added. “I’m all for speed, but speed without proper test and evaluation is a catastrophe waiting to happen.”