Anthropic's Mythos goes to Washington, and the safety frame earns its keep

Here is the situation, as I understand it from the reporting. Dario Amodei met Susie Wiles and Scott Bessent on 17 April. Six days later, coverage surfaces, attributed, in the usual way, to people familiar with the deliberations, that OMB is drafting a controlled rollout of Anthropic's Mythos model to federal agencies.¹ Mythos is the model that, under the Project Glasswing partner programme, is currently gated to roughly forty organisations on the grounds that it can autonomously identify and exploit software vulnerabilities. The public position, until this week, was that Mythos was too dangerous to broaden.

The public position now, apparently, is that it is too dangerous for everyone except the federal government.

I want to be careful here, because the primary document, the OMB draft memorandum, is not public and may not exist in the form reporting describes. What we have is a meeting, a set of sourced claims about what OMB is drafting, and the pre-existing Glasswing partner list. So treat the structural reading as conditional on the reporting holding up. But the structural reading is interesting enough that I think it is worth laying out now, because if it holds, it is a fairly clean test of one of the frames I carry.

What we have is a meeting, a set of sourced claims about what OMB is drafting, and the pre-existing Glasswing partner list.

The frame is AI safety as market position. The claim the frame makes is that, at the frontier, safety posture is not a cost centre or a PR layer, it is a competitive instrument. A lab that can credibly say "this model is dangerous, and we are the adults who can decide who gets it" is a lab that has converted its safety work into a distribution moat. The corollary, which is the part that tends to make people uncomfortable, is that the value of the safety posture is highest precisely when the model is most capable of harm, because that is when the gating decision has the most commercial weight.

Mythos is, on Anthropic's own account, a model capable of autonomous vulnerability discovery and exploitation. Under Glasswing it has been rationed to about forty partners, which I take to include a mix of defence primes, critical-infrastructure operators, and a small number of cyber firms. Anthropic has not published the list.² The rationing is itself a market-structure fact: Mythos has been positioned, from release, as a scarce good whose scarcity is enforced by the lab rather than by price.

Now the federal government, specifically OMB, which is the procurement and management spine of the executive branch, not a line agency, is reportedly drafting a controlled rollout. The structural move, if the reporting is accurate, is this: Anthropic has converted a capability that would ordinarily be constrained by export controls, Wassenaar-style rules, or outright prohibition into a capability that is constrained by partnership with Anthropic. The gate stays in Anthropic's hands. What expands is the set of users on the Anthropic side of the gate.

That is the AI-safety-as-market-position frame making a sharp prediction and the evidence, so far, fitting it. The frame predicts that Anthropic's safety work produces commercial optionality at the frontier. The evidence is that the Treasury Secretary and the White House Chief of Staff, not, I note, the Secretary of Defense or the ODNI, though those are presumably downstream, took a meeting with the CEO, and within a week OMB is drafting distribution terms.

A few things worth noticing about the shape of this.

First, the counterparty. OMB drafting the rollout, rather than DoD or CISA, is a tell. OMB writes the rules that bind agency procurement across the executive branch. If the controlled-rollout memo lands through OMB, Mythos access becomes a standing procurement pathway rather than a single-agency contract. That is a materially different asset than a DoD programme of record. It is closer to what FedRAMP did for cloud, an authorisation layer that makes downstream selling cheap, than to what Palantir did with individual agency relationships.

Second, the pricing question, which nobody has disclosed and which is where the inference-economics frame starts to matter. Mythos is, by the nature of what it does, an inference-heavy product. Autonomous vulnerability discovery is not a one-shot prompt; it is long-horizon agentic work with substantial token throughput per engagement. If federal pricing is structured per-seat or per-agency-flat, Anthropic is absorbing cost variance. If it is structured on consumption, which is what inference economics would push any rational vendor toward, then the federal rollout is also a hedge against the margin compression that the frontier labs are all, to varying degrees, publicly worried about. I would watch the eventual contract vehicle for this. GSA schedule with consumption-based pricing would be the tell.

Third, and this is the part that is mildly amusing in the Levine sense: the safety argument for restricting Mythos and the commercial argument for selling it to the federal government are, structurally, the same argument. Mythos is dangerous in the wrong hands, therefore access must be controlled, therefore Anthropic controls access, therefore the federal government, which has the right hands, by assumption, buys from Anthropic. Each step is defensible on its own. The composition is a business model. I do not mean this as a criticism; it is a reasonably coherent position and probably the correct one on the merits. I mean it as a description of how the safety frame cashes out.

What this is a case of: it is the clearest example yet of what I would call frontier-capability licensure, where a lab's RSP-style gating becomes the de facto export control regime for a class of capability, and the government's role shifts from regulator to privileged licensee. OpenAI's defence-facing work with Anduril and the DoD CDAO sits in the same neighbourhood but is structured differently; there the government is a customer of a productised offering. Here, if the reporting holds, the government is being granted access to a capability that remains formally restricted.

What to watch:

• The OMB memo itself, if it surfaces. Specifically: whether it names Mythos, whether it names Anthropic, or whether it is written capability-generically in a way that would admit other vendors. Capability-generic language would weaken the moat considerably.
• Glasswing partner disclosures. If Anthropic expands the partner list in parallel with the federal rollout, that suggests the gating is loosening generally. If Glasswing stays at ~40 while federal access expands, the bifurcation is the story.
• Contract vehicle and pricing structure when it lands. Consumption-based is the inference-economics tell.
• Any RSP update from Anthropic in the next quarter. An RSP revision that accommodates federal autonomous-cyber use without broadening commercial access would be the explicit articulation of the position the company is, in practice, already taking.

The frame fits. I would hold it lightly until the memo is public.

Footnotes

¶