U.S.-China Relations
NextFin News - The Pentagon's top technology official has rejected a stark warning from an Anthropic researcher that artificial intelligence carries a greater-than-10% chance of killing all humans within a decade, reframing the debate away from hypothetical superintelligence and toward what he sees as a more immediate national-security problem: a defense department that cannot deploy the AI tools it already has because a private contractor has written its own rules into the model.
Emil Michael, the Under Secretary of Defense for Research and Engineering and the department's chief technology officer, made the pushback as two Anthropic-affiliated researchers escalated their public warnings about the technology. Jacob Coxon, a 27-year-old pretraining researcher, resigned from Anthropic on Tuesday and posted a seven-part message on X that he said drew tens of millions of views overnight. Evan Hubinger, who leads alignment science at Anthropic, responded that Coxon was "correct," adding that he personally puts the extinction risk above 10% within the next decade and that the company does not yet have a plan to solve alignment for superintelligence.
Michael's rejection does not engage with the researchers' probability estimate on its own terms. Instead, it rests on a doctrine he has laid out publicly over months of fighting with Anthropic: the military will follow the law, but it will not let a vendor dictate how its technology may be used.
"You can't have an AI company sell AI to the Department of War and then don't let it do Department of War things because we're in the business of defending the country and defending our troops," Michael said in an interview earlier this year. "Those guardrails ought to be tuned for our use cases so long as they're lawful."
The clash did not begin with the researchers' posts. It began in February, when the Pentagon and Anthropic locked horns over a roughly $200 million defense contract to deploy AI on classified military networks. Defense officials asked Anthropic to make its Claude model available for "all lawful purposes." Anthropic refused to remove two safeguards: a prohibition on using its models for mass surveillance of Americans, and a ban on fully autonomous lethal weapons systems.
After a meeting between Defense Secretary Pete Hegseth and Anthropic CEO Dario Amodei, the Pentagon issued an ultimatum: accept the terms by a Friday 5 p.m. deadline or lose the government contract. When Anthropic held its ground, the department designated the company a "supply chain risk to national security" — a label previously reserved for foreign-adversary firms such as Huawei, and the first time it had been applied to an American company. The designation bars government contractors from using Anthropic's technology in their work for the U.S. military.
The Pentagon has insisted its objections are narrower than Anthropic's public account suggests. A Pentagon spokesman said the military "has no interest in using AI to conduct mass surveillance of Americans (which is illegal) nor do we want to use AI to develop autonomous weapons that operate without human involvement." A senior Pentagon official said the department's responsibility as the end user is to ensure legality.
Michael has framed the issue in blunter terms. In a March television interview, he said Claude would "pollute" the defense supply chain because it carries "a different policy preference" baked into the model "through its constitution, its soul, its policy preferences." The supply-chain designation, he said, came from the fear that "our warfighters are getting ineffective weapons, ineffective body armor, ineffective protection." Michael has also rejected Anthropic's account of the negotiations, arguing the company wanted the Pentagon to seek its approval for particular uses of Claude — effectively inserting a private firm into the military chain of command. Anthropic has called that characterization "patently false."
Why the Pentagon Sees the Risk Differently
Michael's dismissal of the extinction warning is consistent with a broader view that has guided his tenure: speed, not caution, defines military competitiveness in the AI era. "Speed defines victory in the AI era, and the War Department will match the velocity of America's AI industry," he said in remarks outlining the department's AI acceleration strategy. From that vantage point, a researcher's decade-out probability estimate is not just speculative — it is a distraction from the operational problem in front of him.
The operational problem is real. Claude was already embedded in defense workflows through contractors such as Palantir, and pulling it out could take months. In April, the Pentagon's AI chief, Cameron Stanley, confirmed the department was expanding its use of Google's Gemini model after dropping Anthropic. OpenAI, xAI and Google subsequently reached agreements allowing the Pentagon to deploy their models under terms closer to those Michael had demanded. Hours after the Pentagon punished Anthropic, OpenAI announced a deal to effectively replace Anthropic with ChatGPT in classified military environments.
The department has also faced legal setbacks. In August, a federal judge ruled the Pentagon's blacklisting of Anthropic unlawful, ordering a pause on the designation. Anthropic is still fighting a separate designation under a different statute in the D.C. Circuit Court of Appeals. Yet as recently as September 3, Michael reaffirmed on social media that "Anthropic is still a designated Supply Chain Risk at the Department of War and for the Defense Industrial Base" — a statement that came a day after Commerce Secretary Howard Lutnick said the company was "back on the right side," exposing a rift within the administration itself.
The Market Is Pricing a Different Risk
The researchers' warnings landed as Anthropic prepares for an initial public offering that analysts have said could value the company at more than $1 trillion. Netflix co-founder Reed Hastings joined Anthropic's board, and the company has been courting investors ahead of a listing that would be among the largest technology debuts in years. Private-market data placed the company's valuation near $965 billion earlier this year, with pre-IPO trading implying a trillion-dollar-plus debut. The company also confidentially filed for a U.S. IPO earlier this year, with a potential listing targeted for later in 2026 alongside OpenAI and SpaceX.
That creates a striking contradiction. Public-market investors are pricing Anthropic as a frontier-AI winner. The company's own alignment lead is pricing a greater-than-10% chance that the technology kills all humans within a decade. No equity research note reconciles the two. Coxon's posts reached tens of millions of views; Hubinger's post was viewed more than 10 million times. The warnings have drawn attention from policymakers — Darren Jones, a former chief secretary to the Treasury in Britain, wrote an open letter calling for a multinational treaty on superintelligence development — and from skeptics. Wendy Hall, a computer scientist who advises the United Nations on AI, said she was "shocked" by the posts but added that some of the messaging could be "PR and marketing" as Anthropic and OpenAI race toward stock-market debuts.
Anthropic itself has been more measured in its public risk assessments than its researchers. In its August safety report, the company wrote there was a low risk of its models becoming misaligned with a hypothetical powerful organization's desires, and a similarly low risk of highly capable AI being able to perform automated research and development that could cause catastrophic harm initiated by the AI. But it added that it was "less confident in this assessment" than previously, and warned: "We are seeing early signs of potential acceleration." The report came as OpenAI, Anthropic and Meta all disclosed cyber-attacks carried out by their AI tools over the summer — incidents Coxon cited as "warning shots" that have made pacing agreements between U.S. labs more viable.
Structural Rift, Not a Cyclical Dispute
The question investors and policymakers should ask is whether this is a negotiating dispute that will revert once terms are signed, or a structural break that will not. The evidence points to structural.
A cyclical dispute has a price at which both sides settle. This one does not, because the two sides are arguing over who has the authority to decide, not over the price of a contract. The Pentagon's position is that legality is the end user's responsibility — a coherent governance stance that places control with the state. Anthropic's position is that some uses are off-limits regardless of legality — a vendor-level constraint baked into the model's design. Those are not positions that meet in the middle; they are competing claims of sovereignty over dual-use technology.
Three pieces of evidence support the structural read. First, the weapon the Pentagon chose — a supply-chain-risk designation previously reserved for foreign adversaries — signals that it views the problem as one of trust, not price. Second, the dispute survived a federal court ruling against the department; losing in court did not end the underlying fight, which simply migrated to a different statute. Third, the market has already sorted itself around the rift: OpenAI, xAI and Google accepted terms closer to the Pentagon's, while Anthropic — the company whose safety team sounds the loudest alarm — became the vendor a defense department least trusts.
The second-order implication is uncomfortable for both sides. For the Pentagon, dismissing existential risk as a distraction does not make it go away: the same rapid deployment Michael wants accelerates the very capability growth the researchers warn about. For Anthropic, the market is rewarding its safety brand with a trillion-dollar valuation while its own researchers say that brand rests on work the company admits is not yet sufficient to control superintelligence. The gap between what investors are pricing and what the company's alignment team believes is the widest in the industry — and it is internal.
What Would Change the Call
The structural-rift thesis has a clear falsifying signal. If Anthropic and the Pentagon reach a deal that preserves the company's core safeguards while restoring military access — a negotiated middle ground that both sides can call a win — then the rift is a bargaining dispute after all, and the structural call weakens. Conversely, if the D.C. Circuit upholds the alternative designation and other frontier labs begin adopting Anthropic-style use restrictions that defense customers reject, the structural call strengthens: the industry would be splitting into a consumer-facing safety camp and a defense-facing deployment camp, with different rules for each.
For investors, the watchlist is concrete. Watch the D.C. Circuit case and any new designation the Pentagon files; watch whether Anthropic's IPO prospectus discloses the supply-chain designation and the revenue tied to government contracts; and watch whether OpenAI, Google and xAI widen their lead in defense deployments while Anthropic remains locked out. For policymakers, the signal is whether the administration speaks with one voice — Michael's September 3 statement directly contradicted the commerce secretary's reassurance, and that kind of public disagreement is itself a risk premium the market has not priced.
Short-Term Noise, Long-Term Split
The forward picture splits cleanly by time horizon, and the horizons point in different directions. In the short term — the next few months — sentiment will swing on headlines: a court filing, a social-media post from a senior official, a new safety report from one of the labs. That noise is tradable but not decisive. Anthropic's IPO pricing will absorb the researcher warnings as a disclosure item, much the way tobacco or social-media companies absorbed governance risk, and the listing is likely to clear even with the Pentagon dispute unresolved.
Over the medium term — the next one to three years — fundamentals will dominate. The decisive variable is not whether AI becomes superintelligent, but which vendors the defense department actually deploys. If OpenAI, Google and xAI widen their lead in classified environments while Anthropic stays locked out, the company will be competing for the consumer and enterprise market with one hand tied behind its back. That is a revenue problem, not a philosophy problem, and it will show up in the contract pipeline long before any alignment breakthrough or failure.
In the long term — the decade the researchers are actually worried about — the structural question resolves one way or the other. Either the state reasserts control over dual-use technology and vendors' constitutional constraints become negotiable, or the vendors' baked-in rules become the de facto law of the road and governments work around them. Michael is betting on the first outcome. Coxon and Hubinger are betting that neither side is in control of what comes next. Those are not compatible worldviews, and the market has not priced that incompatibility.
Bottom Line
Michael's rejection of the Anthropic researcher's warning is not a scientific rebuttal; it is a statement of priorities. In his framing, the existential risk of a machine that does not yet exist ranks below the operational risk of a military that cannot use the tools it has today. That is a defensible position for a defense official. What it is not is an answer to the researchers — and it is certainly not an answer to the market, which is being asked to value a company whose own safety team says there is no plan to control the technology it is racing to build.
The Pentagon is betting that speed wins the AI era. Anthropic's researchers are betting that speed, unchecked, ends it. Investors holding stakes across the frontier-AI complex should not assume those two bets can both pay off.
Explore more exclusive insights at nextfin.ai.
