Anthropic works with White House after AI security scare

Most rules get written before the game starts. You learn them, you play, you live with the calls.

The fight over artificial intelligence in Washington has flipped that order. The rulebook is being drafted after the whistle, by the same two sides that spent the spring trading blows in public.

For two years, the biggest names in AI begged Washington for one thing, a predictable set of federal rules they could plan around. Investors took the hint and priced AI as the surest bet of the decade, pouring money into chips, data centers, and a pipeline of blockbuster public offerings. The working assumption was that government would move at government speed, slow and telegraphed, leaving plenty of room to adjust.

Then one company learned how fast that can change. The slow, telegraphed government everyone planned around showed up overnight, along with a letter and a deadline measured in hours.

That company is Anthropic. Days after the government forced one of its newest models offline, the two sides are now reportedly drafting a joint framework, according to Politico, to judge AI security risks and decide when Washington gets to intervene.

How the Anthropic standoff with Washington began

At 5:21 p.m. EST on June 12, Anthropic received a letter from the U.S. government that it says forced a first for the industry. Within hours, two of its newest models, Fable 5 and Mythos 5, were unavailable.

The order came from the Commerce Department, signed by Secretary Howard Lutnick, and demanded the company cut off all access for foreign nationals, “citing national security,” according to CNN.

More AI:

Anthropic said the letter gave no specifics, but its understanding was that officials had found a way to “jailbreak” Fable 5, slipping past the model’s safety guardrails. The company complied, disabling both models for every customer to stay on the right side of the order.

Anthropic disputes that the risk warranted any of it. Recalling a model used by hundreds of millions over what it called a “narrow potential jailbreak” would, if the same bar were applied industrywide, freeze new launches everywhere, the company argued, according to IAPP.

The trigger came from a rival, by several accounts. Amazon (AMZN) chief executive Andy Jassy raised concerns with Treasury Secretary Scott Bessent that Fable 5’s guardrails could be circumvented, after Amazon researchers coaxed a Mythos-class model into describing restricted cyberattack methods, according to Fortune.

Former administration AI czar David Sacks said officials asked Anthropic to fix or pull the model and issued the export order “reluctantly,” only after the company refused.

None of this landed in a vacuum. President Donald Trump had already directed federal agencies to stop using Anthropic’s products, and the administration has weighed an executive order to strip the company’s models out of government systems.

The export letter was the sharpest cut in a fight that had been building for months.

The standoff did not start in June. Here is how the year built to it:

March 9, 2026: Anthropic sued the Pentagon after being labeled a national security threat, according to federal court records in the Northern District of California.
June 12, 2026: The Commerce Department ordered Anthropic to cut off foreign-national access to Fable 5 and Mythos 5 on national security grounds, CNN reported.
June 13, 2026: Anthropic complied, disabling both models for all customers, according to IAPP.
June 18, 2026: The White House and Anthropic reportedly began drafting a joint risk framework, Politico noted.

The White House and Anthropic are now reportedly building a shared framework to assess AI security risks.

sankai / Getty Images

What the new White House framework means for AI investors

Here is the turn that matters. The White House and Anthropic are now reportedly building a shared framework to assess AI security risks and guide future intervention, according to Politico.

Read plainly, that is Washington and a leading lab co-writing the conditions under which the government can reach into a private company and switch a product off.

When I lined up the dates, the sequence is what jumped out at me. The government did not wait for a rulebook. It used export-control powers, the kind normally aimed at chips and weapons headed to adversaries, to pull a finished piece of commercial software offline first, and is now writing the standard second.

That ordering is the whole story for anyone holding AI exposure.

Investors cannot buy Anthropic directly. It is privately held and only now preparing an initial public offering (IPO), having hired law firm Wilson Sonsini for work that could begin as early as 2026, according to TheStreet. The precedent, though, lands on every name they can buy.

Nvidia (NVDA), the closest liquid proxy for the entire trade, holds stakes in both Anthropic and OpenAI. The IPO pipeline investors are counting on, from Anthropic to OpenAI, now carries a risk that did not exist before June 12, 2026.

A regulator can decide a finished, shipping model is a national security problem and act before a court weighs in.

What strikes me is how cleanly this transfers. If a model used by hundreds of millions can go dark overnight over a single disputed jailbreak, that same switch hangs over the data-center buildout, the chip orders behind it, and the queue of AI offerings investors are waiting on.

The framework being drafted now decides whether June 12 was a one-time flare-up or a permanent lever.

The more complex problem is that the government’s reported demand may be unmeetable. Officials want Anthropic to test its own models, patch every jailbreak, and report findings before release, an effectively zero-defect standard.

Anthropic and independent security researchers say that is not how these systems work, since a model’s outputs are probabilistic and no guardrail holds against every possible prompt.

Where the Anthropic fight goes from here

The surprising part is not that the two sides are talking. It is who holds the pen. A framework written jointly by the regulator and the most targeted lab will carry both Anthropic’s read of what is technically possible and the administration’s appetite for control, and it will likely become the yardstick by which every other AI company gets measured.

Allied governments are already asking how reliable American AI can be if Washington can switch it off overnight, a question that now trails every export deal in the sector.

For investors, the signal is simple and a little uncomfortable.

The biggest risk to the AI trade in 2026 may not be a demand miss or a circuit overheating in a data center. It may be a letter, signed on a Friday, that no earnings model can price.

Watch what this framework actually says about triggers and timelines. That language, more than any chip roadmap, will tell you how much of the AI boom Washington intends to keep within reach.