Anthropic Locks Down Mythos as White House Sharpens Knives

Anthropic Ships a Weapon, Then Takes It Back

Anthropic released Mythos, a tool that finds software vulnerabilities, and immediately restricted it to roughly 200 partner organizations after concluding broader access was too dangerous. This is the most honest thing a frontier lab has done this quarter, and also the most damning. The pitch for years has been that capable models can be deployed safely with the right guardrails. Mythos is the counterexample the labs themselves are now conceding exists. A model good enough to harden your code is, by construction, good enough to break someone else's, and Anthropic decided the public internet did not need that capability democratized.

The gated rollout will reassure exactly nobody. Two hundred partners is still a sprawling attack surface, the weights presumably exist on Anthropic infrastructure that someone will eventually try to exfiltrate, and competitors are already training similar capabilities with less compunction. The honest read is that we now have a public admission that frontier offensive cyber is here, and the only thing standing between it and the open web is a partner agreement.

The White House Picks Its Lab

While Anthropic was making the most safety-conscious release decision of the year, the Trump administration was escalating a campaign against the company. The framing from the West Wing is that Anthropic's safety posture is slowing American AI dominance against China. The framing from anyone who reads the news is that Anthropic's CEO has been politically inconvenient and the administration has decided the cost of punishing him is low.

This is the precedent that matters. A lab that voluntarily restricts a dangerous tool gets pressured by its own government for being insufficiently aggressive. Every other frontier lab is watching, doing the math, and concluding that conspicuous caution is now a political liability. The selection pressure is pointed the wrong direction, and it is pointed there on purpose.

The UK Cancels Adolescence

The UK announced it will ban children under 16 from social media, following Australia's lead. The BBC, to its credit, immediately noted that nobody has demonstrated this will actually fix anything. Age verification regimes leak, kids route around them, and the underlying product, infinite-scroll engagement optimization, remains legal and lucrative for everyone over the cutoff.

It is still probably the right call. The platforms had fifteen years to self-regulate and produced a teen mental health crisis instead. If the policy mostly fails, it fails in the direction of less algorithmic feed time for thirteen-year-olds, which is a tolerable failure mode. The interesting question is whether the same logic gets applied to AI companions in eighteen months, when the harm signal is louder and the lobbying is better funded.

Meta Apologizes to Itself

Meta CTO Andrew Bosworth told staff the AI reorganization was catastrophically mismanaged and promised stability, better communication, and restored perks. This is what happens when you spend a reported nine figures per hire poaching a superintelligence team and then forget to tell the existing org what its new reporting lines are. Bosworth's mea culpa is unusual mostly because it is internal honesty leaking out, not because the underlying chaos is unusual. Every hyperscaler is running a similar reorg right now. Meta is just the one apologizing.

Meta also quietly turned its new AI search Mode loose on users' public posts as training and retrieval substrate. Public has always been the magic word that means whatever the platform needs it to mean, and this is the latest iteration. If you posted a status update in 2011 thinking it would scroll into the void, surprise, it is now grounding context for a generative search product.

SpaceX Cashes In During the Fight

SpaceX shares jumped 19.6% on their second trading day after a record IPO, handing Musk a fresh war chest at the exact moment the administration is rearranging the AI policy landscape in ways that favor his other companies. The market is not pricing rockets. It is pricing political access, Starlink's distribution moat, and xAI's claim on the compute and capital that an unfriendly Anthropic environment frees up. None of these stories are independent. Anthropic gets squeezed, Meta loses the plot, the UK fences off the kids, and the guy with the rockets ends the week richer. That is the slot.