clening's Shortform

clening

This is a special post for quick takes by clening. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

Mythos, Leverage, and Tech Extensity

In late February, the Trump administration and the Department of Defense took the unprecedented step of declaring Anthropic a supply-chain risk, and cutting the company off from government contracts. I predicted in March that despite the administration’s efforts, Anthropic would not only prevail, but would come out ahead:

Bluntly, Trump needs Anthropic (and by extension, the hyperscalers, VC firms, and sovereign wealth funds [who support Anthropic]) far more than they need him. Even if the courts do side with the administration, Trump still has to contend with all the essential companies who have funded his largesse, and the fact that so many of these donors have a vested interest in Anthropic succeeding. That’s extensity.

No matter how powerful Trump thinks he is, no dictator can afford to piss off the coalition that actually keeps him in power. So, in the end, I don’t think Trump’s theatrics will hurt Anthropic much, because there are far too many big, important essential players who need Anthropic, and who have the leverage and ability to get Trump and Hegseth to blink first.

I mention all this, because on Friday, Dario Amodei met with senior White House officials, in an effort to reach a detente. I thought that was interesting, but what I was missing was the specific leverage point: Claude Mythos. Matija Vidmar explained things nicely.

Suddenly, a few things clicked.

Based on their Project Glasswing announcement on April 7, Anthropic has almost certainly been working on its Mythos frontier AI model for months now, and have been testing it before the spat with the administration reached its high-drama phase.
Anthropic believes their model is powerful enough that it needs further evaluation prior to release.^[1] Rather than keep that research in-house, they decided to selectively release the model to a handful of security companies (all of whom took stock hits on the April 7 announcement), the Linux Foundation, and tech firms.
Those tech firms, which include AWS, Google, Broadcom, Microsoft, and Nvidia—all have a vested interest in Anthropic’s success. JPMorganChase’s inclusion also isn’t coincidental. Goldman Sachs and Citigroup are also purportedly testing Mythos.
While these companies might stay quiet publicly (to avoid a predictable administration meltdown), it’s near certain that every last one of these companies was fiercely advocating on behalf of Anthropic internally. It’s no coincidence that JD Vance and Scott Bessent met with Anthropic, Google, xAI, and Microsoft a day or so prior to the Mythos drop. A separate meeting between Scott Bessent, Jay Powell and the Wall Street banks occurred a few days later.
I believe Amodei pushed back on the administration’s demands precisely because he knew Mythos presented a risk, and that the administration’s theatrics could only go so far. As I wrote, “Amodei can afford to take a principled stance against the administration not necessarily because the company is braver or more patriotic than other AI companies, but because Amodei knows that Anthropic has extensive reach and leverage.”
It would be politically suicidal for the administration not to back down at this point. Tanking Anthropic because Amodei didn’t kiss the ring sufficiently risks alienating Trump’s coalition (which he can’t afford to lose). Icing out Anthropic from government contracts also creates a genuine national security risk. If CISA or the DOD, or NSA can’t use Mythos to harden their systems, those systems are exposed to threat actors and nation states who are happy to use Anthropic models.
There’s literally no upside for the administration to play a game they can’t win, and I think Anthropic knows this. I suspect that even if Trump doesn’t get it, JD Vance, Scott Bessent, Susie Wiles do.

My current forecast on the supply chain risk question

On March 22, I used FeedForward (my internal research & forecasting engine) to run a forecast on whether Anthropic will still be designated a supply chain risk by the US Government on May 31, 2026.^[2] Here’s the current assessment, with increasing/decreasing likelihood measures.

Current Assessment

Confidence: 5% (updated 2026-04-17)

Key Factors

Increasing likelihood:

Mythos creating normalization pressure - Treasury Secretary Bessent’s characterization of Mythos as strategically important, combined with White House plans for agency access and Treasury CIO seeking Mythos access, suggests the government may need to modify or lift the designation to utilize the technology for national security purposes
Pentagon’s practical needs - Wall Street CEOs summoned by Bessent and Powell to discuss Mythos cybersecurity applications, with major banks (JPMorgan, Goldman Sachs, Citigroup) reportedly testing it, indicates government is prioritizing defensive cybersecurity capabilities over the ban
Expedited litigation timeline - May 19 oral arguments in the DC Circuit supply-chain risk case, create possibility of ruling before May 31, though extremely tight
Political optics of inconsistency - Government simultaneously praising Anthropic’s strategic value while maintaining supply chain risk label creates unsustainable tension

Decreasing likelihood:

Appeals Court definitively ruled against emergency relief - DC Circuit panel refused to block the supply-chain risk designation, with explicit reasoning citing “judicial management of how, and through whom, the Department of War secures vital AI technology during an active military conflict”
Court split confirms prolonged litigation - San Francisco judge found DOD likely acted in bad faith and ordered removal, but DC Circuit’s opposite ruling means resolution requires higher court intervention, extending timeline
Trump-appointed judges’ reasoning - The DC panel (including Gregory Katsas, former Trump deputy counsel, and Neomi Rao from OMB) emphasized national security deference and military conflict context, suggesting even favorable procedural arguments won’t overcome security claims
No evidence of settlement negotiations - Despite the normalization signals, no articles indicate active DOD-Anthropic settlement discussions.

In short, what we’re witnessing is extensity in action. It’s not just a question of what AI model some agency uses for its internal chatbot. It’s about what happens when a single company becomes so integral that it can push back against the traditional levers and forces (regulatory, political, economic, or otherwise) that govern behavior.

^{^}
Though some, like Devansh have called out the media for blindly parroting Anthropic’s press release and not, the primary sources, CVEs, exploit code, etc. He dug into those details and suggested that many of the capabilities might just be hype.
^{^}
I ran a separate forecast on whether they would still be kept out of government contracts, notwithstanding the supply chain risk designation. I have much higher odds (43% chance vs. the community’s 75% chance) of that being the case, but even those odds are decreasing.

I believe Amodei pushed back on the administration’s demands precisely because he knew Mythos presented a risk

From what's publically known it seemed, they had a deal that prevented the administration to use Claude as a cyberweapon. The administration asked to use it for all purposes.

Amodei's red lines weren't about using Mythos as a cyberweapon either to hack targets or spread disinformation. He did not push back on either of those. He only pushed back on the model being used to make unsupervised kill decision and domestic surveillance.

To the extend that Mythos has dangerous capabilities it's probably in the cyberwar front and Amodei failed to push back on that front.

FWIW, I covered more of Anthropic's boundaries (which, if I'm reading you correctly, weren't actually all that substantial) in this post. I didn't rehash that here.

But it's possible he has other boundaries, e.g., hostile nation-state acts, cyberweapons, but those were not explicitly demanded by the DOD/administration. I don't know. I wasn't in the room.

FWIW, I covered more of Anthropic's boundaries

Given that you didn't talk about Anthropic's boundaries that they have in their actual contract, I think you did a poor job at that. Their Exceptions to our Usage Policy along with public statements by Dario suggests that the current contract has explicit limits to prohibiting use for disinformation campaigns, the design or use of weapons, censorship, domestic surveillance, and malicious cyber operations.

Anthropic was one of several AI companies to offer its AI tools at the bargain rate of $1 annually per agency, via the GSA’s 2025 OneGov initiative.

This looks like a misunderstanding of the dynamic. The DOD has a 200$ million contract with Anthropic.

Also, there’s a reason the DOD and members of the intelligence community use Claude and not Grok for analysis and warfighting capabilities, and it’s not just the GSA deal. Claude is a superior product for many use-cases, which is why Claude was the only LLM designated to handle classified materials until the 28th of February.

Given that the DOD does have a contract with xAI/Grok that's again misleading. It also suggest to a naive reader that the Claude models being superior is the reason, when the fact that Claude running within the classified AWS enviroment when Grok doesn't is probably more significant.

But it's possible he has other boundaries, e.g., hostile nation-state acts, cyberweapons, but those were not explicitly demanded by the DOD/administration.

The administration demanded no specific usecases explicitly. They demanded that all lawful use is allowed in the contract. I don't think you need to have been in the room to know that the position of the administration was "all lawful use". Hegseth seemed quite clear to me that he considers all the explicit limits in the current contract bad. This means getting rid of the current explicit limits for disinformation campaigns/malicious cyber operations.

With respect, I'm not sure you fully read the post. I called out the existing contract specifically, though I did not mention the dollar amount.

On February 26, after weeks of negotiation, talks broke down between the Pentagon and Anthropic after Anthropic’s CEO, Dario Amodei refused to accede to DOD demands for unfettered “lawful”1 uses of its AI tools by the military. Anthropic was generally fine with Claude being used by the military generally (as it had been since July 2025), and by other strategic military contractors, such as Palantir, Amazon, Oracle, and Lockheed Martin, for everything from supply chain logistics and cyber operations, to foreign surveillance & intelligence gathering.

I also mentioned the fact that Anthropic was unique in that it ran in the classified environment. I quoted Anthropic's language explicitly.

Given that, I don't think we're actually in disagreement here, unless you're objecting to my tone or something not related to the substance of my piece, in which case, there's nothing more I have to say.