The Week Washington Decided Which AI Companies Live

The Policy Is Whoever Trump Likes This Week

The defining story of the week is not a model release. It is the quiet confirmation that American AI policy is now a personal favor system administered out of the West Wing, with no statute, no rule text, and no appeal.

Anthropic spent the week being slowly strangled. The White House ordered the company to revoke SK Telecom's access to Claude, citing China ties that Anthropic was apparently supposed to have inferred from a memo nobody wrote. Then the administration ordered foreign national access revoked across the board, which had the practical effect of disabling Fable 5 and Mythos 5, Anthropic's newest models, because the company could not certify compliance with rules that do not exist in publishable form. Reporters at Wired and NPR spent the week asking the obvious question, which is what rule, exactly, Anthropic had broken. Nobody in the administration could answer. Trump himself went on record saying Anthropic posed no security threat, while his own agencies continued throttling the company in real time.

This is not regulation. Regulation is written down. This is mood.

Anthropic staff are reportedly furious, which is the correct response and also a completely useless one. The Financial Times noted, with the gentle cruelty that paper does best, that Anthropic spent the year warning more loudly than OpenAI about AI dangers, and is now being punished more loudly than OpenAI for it. The lesson the rest of the industry is absorbing is not subtle. Safety advocacy is a liability. Quiet compliance is a survival strategy. Anthropic is the cautionary tale the administration wants every other lab to study.

Meanwhile, The Favored Son

While Anthropic was being disassembled by inference, Elon Musk had what can only be described as a structurally excellent week.

The Department of Justice invoked national security to kill a pollution lawsuit against a SpaceX datacenter, citing the company's involvement in the ongoing Iran conflict. Read that sentence again. A private datacenter operated by the world's richest man received federal litigation immunity because his rockets are doing geopolitics. The environmental claims did not get adjudicated. They got classified out of existence.

SpaceX then exercised its option to acquire Cursor for sixty billion dollars in an all-stock deal, folding the coding assistant into what Charlie Warzel accurately described as a seven-headed hydra of finance. SpaceX is no longer a rocket company. It is a holding structure with launch capability attached. Musk is also pushing a SpaceX-Tesla merger that legal experts agree shareholders have limited ability to block, because the governance documents Musk wrote for Musk turn out to favor Musk.

The contrast with the Anthropic treatment is the entire story of the week. One company gets disabled by unwritten rules invoked through a press leak. Another gets federal lawsuits dismissed and acquires a major AI startup with the casual energy of a man ordering lunch. Both outcomes flow from the same office. The difference is loyalty.

And the Tesla on Autopilot that crashed into a Harris County home this week, killing a woman inside her own house, will not generate the same urgency. The DOJ is not going to invoke national security to shield her family's claim, and it is not going to invoke it to block one either. She will be a settlement, eventually, and a footnote.

The Tools Are Now Too Dangerous For The Public

Anthropic also did something this week that deserves more attention than the political theater around it. The company restricted its Mythos vulnerability-finding tool to 200 partner organizations, declining to make it generally available. The stated reason is that a tool which autonomously discovers software vulnerabilities at scale is, in unfriendly hands, a weapon.

This is correct. It is also a milestone. We have crossed into the territory where frontier labs are explicitly conceding that some of their products cannot be sold to the public without producing catastrophic externalities. The vulnerability-finding capability is real, it works, and the only thing standing between it and a thousand ransomware crews is an access list maintained by a company the White House is currently trying to break.

If that sentence does not unsettle you, read it again.

The credential breach this week is the companion piece. Thousands of sensitive networks, including Oracle, Lenovo, FedEx, and NATO contractors, had login credentials exposed in a single dump. The defenders are losing the basic blocking and tackling of credential hygiene. The attackers are about to get Mythos-class tools, if not from Anthropic then from whoever copies the paper next quarter. The Atlantic ran a piece this week featuring a former cancer detection researcher who said she would rather take her chances with undetected cancer than live through the pace of AI development she is watching. That is not a quote that ages well in either direction.

Militaries Are Now Taking Advice From The Machines

MIT Tech Review reported, with admirable understatement, that armed forces around the world are integrating AI models into command structures as decision-making advisors. Not targeting systems. Not logistics optimizers. Advisors. Models that recommend what to do.

The failure modes here are not speculative. We have spent three years documenting that frontier models hallucinate confidently, exhibit sycophancy toward the operator's apparent preferences, and produce coherent reasoning chains for conclusions that are simply wrong. Now imagine that capability deployed next to a brigadier who is tired, who has been awake for thirty hours, and who would like a second opinion that agrees with him.

This is the deployment context safety researchers have been screaming about for a decade. It arrived this week with no fanfare, because the news cycle was busy with Musk acquisitions and Anthropic's slow execution. The militaries did not wait for the public to be ready. They never do.

The Jobs Story Stops Being A Forecast

The labor news this week shifted register. For two years the AI jobs story has been a forecast, a McKinsey chart, a Davos panel. This week it became a schedule.

Accenture's stock hit a seven-year low because investors finally priced in what consultants do for a living and noticed a language model can do most of it. Lloyds hired 300 AI experts and told staff, in roughly these words, that the hires now mean the cuts later. UnitedHealth disclosed billions in AI investment to deploy bots that call doctors, read charts, monitor calls, and handle claims, which is to say, the entire administrative middle of American healthcare. If you have ever spent ninety minutes on hold with an insurance company, the good news is the hold music is about to get more efficient. The bad news is everything else.

The single most clarifying jobs story of the week came from an AI startup that is dispatching free human cleaners to New York apartments. The cleaners clean. They are also wearing sensors and cameras, gathering training data for the cleaning robots that will replace them. The workers know. They took the job anyway, because rent is due now and the robots are due later. There is no more honest illustration of the current labor compact than this. You are paid, today, to teach the system that will not need you tomorrow.

Amazon, for its part, spent the week retaliating against workers who testified in Seattle hearings in favor of datacenter regulations. Meta's CTO admitted the company's AI reorganization was catastrophically mismanaged and promised perks would return, which is the corporate equivalent of bringing flowers home after the fight. Morale in Meta's AI division is reportedly in the basement. Waymo's national rollout is stalling on political opposition rather than technical limits, which is genuinely new. The technology is ready. The cities are not.

The Data Was Never Yours

The Verge mapped 21 million songs that have been ingested as training data, a database you can search to find your own work. Meta launched an AI search mode that trains on your public Facebook posts, by which they mean every post you ever made under a privacy regime that no longer applies. The Guardian documented brands using undisclosed AI influencers to pose as real customers, a practice that is illegal in spirit, legal in fact, and ubiquitous in execution.

The consent framework for the training data era has fully collapsed and nobody is rebuilding it. Jensen Huang spent the week saying, with the relaxed honesty of a man who has already won, that society will need to fundamentally change to accommodate AI. He is correct, and the change he means is that you adapt to the model, not the other way around. Peter Thiel's secret club, whose membership files leaked this week, rates its members by wealth and influence on an explicit scale. The people building the future are not coy about how they rank the rest of us.

What To Watch Next Week

Three threads to track.

First, whether Anthropic gets a written rule or stays in regulatory purgatory. A written rule is survivable. Purgatory is not, because purgatory means every model launch needs informal White House blessing, and informal blessings can be revoked at the speed of a Truth Social post. If no rule appears by Friday, assume the absence is the policy.

Second, whether any other lab adjusts its public safety posture in response to the Anthropic treatment. OpenAI has been quieter on safety this year, and the FT noticed. If Google DeepMind or Meta dial back their safety communications next week, the signal will be unmistakable. Speaking honestly about risk is now a regulatory liability under this administration.

Third, the Tesla Autopilot fatality in Texas. The pattern with these cases is a quick NHTSA acknowledgment, a slow investigation, and a settled lawsuit two years later. Watch whether the DOJ's new theory of national security litigation immunity, deployed this week for the SpaceX datacenter, gets extended to Tesla product liability. If it does, the legal architecture of American consumer protection against Musk-affiliated products is functionally over.

The average doom score this week was 4.1, which feels low for a week in which the executive branch demonstrated it can disable a frontier lab by phone call. The number is low because most of the individual stories are mundane. The configuration is what is dangerous. A favored champion, an out-of-favor competitor, an unwritten rulebook, military deployment, and a labor market quietly training its own replacements. None of those is a five out of ten on its own. Together they are the week the structure clicked into place.

See you next week. Bring water.