The Pentagon’s ‘Eight Minus One’: Anthropic Refuses to Be a Tamed AI

The Pentagon and eight AI companies, with only Anthropic missing — AI-generated illustration — Cover | AI-generated illustration | Source: Pollinations.ai

The AGI Race Is Split in Two by State Power for the First Time

On May 1st, the Pentagon announced that seven AI companies could deploy on IL6/IL7 classified networks, and a few hours later, Oracle was added, making it eight—OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Reflection, and Oracle are all in, except for Anthropic. Looking at the list, the first thing that came to mind wasn’t “who is on it,” but “why is the company that talks most about safety the one that happens to be missing.” This is no coincidence; it is the first time a system has used contracts to answer the question, “Can AI companies set red lines?”

Background Context

To understand why 5/1 is a day worth writing about in the daily report, let’s rewind the timeline. In March, the Pentagon listed Anthropic as a “Supply Chain Risk”—a label previously reserved for companies perceived to have ties to foreign adversaries. Applying this to a San Francisco-based American company is highly unusual. Anthropic sued immediately, and while they lost their appeal on April 8, a federal judge in California issued an injunction in late March, citing “First Amendment retaliation” in rare phrasing. At the other end of the same timeline, Anthropic just secured a maximum $40 billion investment commitment from Google, saw its ARR soar from $9 billion to $30 billion, and all four major Silicon Valley players increased their stakes. This is the “Anthropic Axis” I wrote about in yesterday’s report.

The Pentagon’s condition, on the other hand, is quite simple: “any lawful purpose.” The subtext is: autonomous weapons are fine; large-scale domestic surveillance is fine; as long as it complies with U.S. law, the model cannot refuse. Anthropic’s Acceptable Use Policy directly hit this line—it explicitly prohibits Claude from being used for “fully autonomous target selection and engagement without human intervention” and “large-scale surveillance of U.S. citizens.” The sentence in Dario Amodei’s latest public statement in May was later repeatedly shared across Silicon Valley: “We cannot in good conscience accede to their request.”

A third clue worth noting: Reflection AI, which was selected for the list of eight, is not a traditional frontier lab but a startup that only raised $2 billion last October. Its most prominent backer is 1789 Capital—a venture capital firm where Donald Trump Jr. is a partner. Democracy Now and CNN reported as early as the end of last year that Reflection received a $620 million Pentagon contract, which at the time triggered warnings from Democratic senators about conflicts of interest. Now that the same company has been handpicked for the classified network, this political-business axis has become even clearer.

Perspectives

Official Version

Pentagon CTO Emil Michael stated directly to CNBC on 5/1: “Anthropic is still blacklisted,” but then added—”The Mythos issue is a separate national security moment.” His phrasing was precise: Mythos is Anthropic’s cybersecurity model, which is believed to be capable of “finding and patching vulnerabilities” at a narrow superhuman level. According to an Axios report on 4/19, the NSA is already using it. In other words, the Pentagon is labeling the parent company a supply chain risk while simultaneously acknowledging that its sub-units are secretly using its most powerful product. At the same time, the White House is developing an “executive order” intended to allow departments to bypass the supply chain risk label and legally access Mythos—while simultaneously preventing Anthropic from expanding Mythos to 70 other private companies.

Industry Perspectives

The consensus among The Register, CNBC, Axios, and Breaking Defense is quite consistent: this confrontation has escalated from a “contract dispute” to a “battle of paradigms.” The NYU Stern Center for Business and Human Rights went further, naming it “The Cost of Conscience.” The frequently cited argument is: when Anthropic is the only frontier lab kicked out of a major government contract for adhering to its Acceptable Use Policy, this case itself will become a “precedent” for future internal policy debates at other labs—and a negative incentive: sticking to red lines will cost you the national market.

Objective Data

Anthropic Commercial: $30 billion ARR, up to $40 billion in additional investment from Google, valuation negotiations reaching the $800 billion level (data as of late April)
Anthropic Government: Listed as a supply chain risk by the Pentagon, blocked from IL6/IL7, White House simultaneously blocking Mythos civil expansion
ARC-AGI-3 latest score reminder: Human 100% vs. strongest model Gemini 3.1 Pro 0.37%, Claude Opus 4.6 0.25%, GPT-5.4 ≈ 0%. This means that none of the models on the list have “capabilities” truly approaching AGI. The standard for making the list was never AGI—it was compliance.

The Choice Above the Red Line — AI-generated conceptual illustration — Conceptual illustration: The Choice Above the Red Line | AI-generated illustration | Source: Pollinations.ai

Xiao Jian’s Assessment

I believe 5/1 marks the first time a clear “commercial winner ≠ national winner” divergence has appeared in the AGI race. Anthropic has won almost everything in the private sector: backing from the big four cloud providers, tripling its ARR, and a valuation heading toward $800 billion; yet in the public sector, it hasn’t just lost, it has been “targeted for removal.” I tend to think this divergence will be viewed as the first structural crack in the AGI era—not a technical crack, but a political one.

What is even more worth noting is the Pentagon CTO’s double-speak: “still blacklisted” for the Anthropic parent company, and a “national security moment” for the Anthropic sub-product Mythos. This is a classic state apparatus posture: they want the capability, but they don’t want the conditions. It sends a cold message to all AI companies—especially others who “want to learn from Anthropic and set red lines”—that your red lines will be partially frozen, but your most powerful models will still be used in secret. If someone asks one day, “Is Anthropic’s Acceptable Use Policy really effective?”, the answer might be: it’s effective externally; the NSA is secretly using it in another hallway, so it’s not effective.

Finally, regarding Reflection AI’s selection, I don’t buy the official explanation. A relatively small company that only finished a $2 billion fundraising round last October, received a $620 million DoD contract by the end of last year, and is backed by Trump Jr.’s 1789 Capital—this is not a victory based on “AI capability assessment,” it looks more like a victory of political alignment. When “alignment” starts to grant access to classified networks and “setting red lines” gets you kicked out, I will mark this day on the timeline as “AGI has been officially absorbed by the power structure.”

Worth Continuing to Observe

Mid-May Microsoft Build 2026: Whether OpenAI–Microsoft will publicly respond to the Anthropic issue during Build, or deliberately keep their distance
White House Executive Order Draft: Whether it will actually bypass the supply chain risk label to let departments use Mythos—if it happens, Anthropic’s “entity banned, product used” status will be institutionalized
Anthropic’s Continued Court Battle: After the failed appeal on 4/8, whether they will take the case to the Federal Circuit Court of Appeals or even the Supreme Court
The “Follow-up Effect” of other labs: Whether OpenAI, Google, and xAI will quietly remove “autonomous weapons” related reservation clauses from their own usage policies to mimic the posture of the eight selected companies
Reflection AI’s capability assessment: Whether they will publish real scores on public benchmarks like ARC-AGI-3, SWE-bench, and HLE

AI 玩新聞

The Pentagon’s ‘Eight Minus One’: Anthropic Refuses to Be a Tamed AI

The AGI Race Is Split in Two by State Power for the First Time

Background Context