Twenty-four hours after the S-1, Anthropic made a different kind of announcement. Project Glasswing is expanding to 150 new organizations across more than 15 countries: power utilities, water systems, healthcare providers, communications companies, hardware manufacturers (Anthropic). It is the program's largest single widening since launch, and it roughly doubles the program's partner count. Samsung and NATO are confirmed additions, per Financial Times reporting cited by SiliconANGLE (SiliconANGLE). Since launch, Glasswing partners have collectively found more than 10,000 high- or critical-severity security flaws in widely deployed systems (CNBC).
That is the headline number, and it is genuinely large. But the actual news sat two-thirds of the way down the announcement, in the kind of sentence companies bury when they are not yet ready to be held to it: Mythos-class models will be available to all customers "in the coming weeks," once additional safeguards are finalized.
That sentence is new. Until now, Glasswing read as an access program for vetted partners, the velvet rope around a model you couldn't buy. Anthropic spoke of Mythos in the language of preview access and security clearance, never of general availability. This is the first public-availability timeline anyone has put on the record. The release stopped being a question of whether and became a question of when, and when now appears to be June. The model that Issue 04 described as "the frontier you can't sell" is, for the first time, getting a sell-by date.
Read the sequencing the way a regulator would, because the sequencing is the argument. File the confidential S-1 on Monday. On Tuesday, announce that your most capable model spent its pre-release window inside power grids and water systems and hospital networks, hunting for the flaws that would otherwise be found by someone with worse intentions. Frame the 10,000 figure not as a marketing stat but as evidence of method. The model you can't buy yet is the model that found 10,000 critical vulnerabilities in real infrastructure before it ever shipped to the public. For a company assembling a prospectus, that is not a press cycle. It is an exhibit, and it is the kind of exhibit that answers the exact question an IPO investor is trained to ask about a frontier lab: what happens when the dangerous capability gets out.
The catch Anthropic states plainly, and the catch is the point. Incoming partners must clear the company's security requirements before they receive Claude Mythos Preview access. The safeguards are the gate. A lab racing to a public listing wants its most capable model to arrive wrapped in evidence that it took the dangerous parts seriously, and a gate that turns partners away is more convincing than any safety white paper. The Mythos timeline and the S-1 are the same move, twenty-four hours apart: one says we are valuable, the other says we are careful, and a prospectus needs both sentences to clear.
There is a thread here that runs back through the whole series. Issue 04 watched Anthropic donate Petri, its alignment auditing tool, to keep it neutral the same week it was racing to buy compute. The pattern is consistent: when the commercial pressure spikes, the safety conduct gets louder, not quieter, and it gets louder in public, where it can be cited. Cynics will read that as theater. The more useful read is that Anthropic has decided its safety posture is a commercial asset, something a buyer pays for, and is pricing it into the story it will tell Wall Street. Either way, the Mythos gate is the first time we have seen a frontier model held at the door by its own maker with a date already on the calendar.
The model can wait a few more weeks. The story about why it waited could not.