On the morning of February 28 Beijing Time, Sam Altman posted a tweet: "Tonight, we reached an agreement with the U.S. Department of the Army to deploy our models on their classified network."
About twelve hours earlier, at 8:00 PM Beijing time on February 27, the same person sat in front of the CNBC Squawk Box camera and calmly said: “Regarding Anthropic, despite many disagreements between us, I generally trust this company. I believe they truly care about safety.” He also said: “I don’t think the Pentagon should threaten these companies with the Defense Production Act.”
Less than twelve hours apart, two statements from the same mouth. What happened in between is worth explaining carefully.
Two similar terms, two different outcomes
Place the core contents of the two contracts side by side.
Anthropic's requirements: Claude must not be used for mass surveillance of U.S. citizens, nor for autonomous weapon systems without human involvement.
When Altman announced the protocol in a tweet, he cited two identical principles: “We have long believed that AI should not be used for mass surveillance or autonomous lethal weapons, and that humans should always be in the loop for high-risk automated decisions.” He also wrote: “The Pentagon agrees with these principles, has incorporated them into law and policy, and we have written them into our contracts.”
The text is almost identical.
One was banned, labeled a “supply chain risk,” and personally called an “extreme left-wing woke company” by Trump on Truth Social. The other secured a contract and gained access to the Pentagon’s classified network—Sam Altman used the phrase “reached an agreement” in a tweet—calm, businesslike, as if a routine B2B transaction had been completed.
Why are the results completely different for similar clauses?
The answer is not in the terms, but in the logic behind them.
There is a key background fact to clarify first: Anthropic is the only one of these four companies (the other three being OpenAI, Google, and xAI) that has been permitted to connect its AI to the Pentagon’s classified network. OpenAI’s original contract covered only non-classified, everyday office scenarios. This negotiation essentially involves OpenAI seeking access to the classified network, with the Pentagon’s condition for entry being the controversial “all lawful purposes” clause. Anthropic is already inside, but the Pentagon is now demanding that it remove the security gate it set up upon entry.
The Pentagon doesn't care what is written, but who calls the shots.
To understand this, you first need to understand what Dario Amodei of Anthropic actually said in that open letter.
He wrote: "Anthropic understands that military decisions are made by the Pentagon, not by private companies. We have never objected to specific military operations, nor have we attempted to temporarily restrict the use of our technology."
Then he changed the subject: "But in rare cases, we believe AI could undermine rather than uphold democratic values. The threats will not change our stance: we are at peace with our conscience and cannot accept their demands."
Anthropic requires that the principles be codified into the contract terms as binding obligations; if the other party violates them, Anthropic reserves the right to refuse further services.
What did the Pentagon hear? A private company telling the government's military: In certain situations, I will not carry out your orders—the boundaries are set by me.
This is unacceptable for any military. It’s not because they truly want mass surveillance, but because the very question of “who gets to decide” strikes at the most sensitive nerve in military command structures. Jerry McGinn, head of military procurement, put it bluntly: defense contractors typically have no authority to tell the Pentagon how their products can or cannot be used—“otherwise, every contract would require negotiating specific use cases, which isn’t practical.”
OpenAI gave a completely different answer.
Altman told employees in a memo that OpenAI will propose to the Pentagon building its own “safety stack”—a multi-layered protection system composed of technical controls, policy frameworks, and human review, embedded between AI models and their real-world use. OpenAI also indicated it could station cleared researchers within secure networks to continuously monitor AI behavior; models would be deployed only in the cloud and would not be integrated into edge systems such as drones.
You come, you oversee, you see everything that happens. If something goes wrong, we share the responsibility—not you coming to me for an explanation.
“Rules are hardcoded, and I execute” and “I embed, and you supervise” represent two entirely different power dynamics, and the Pentagon accepts only the latter.
What OpenAI is best at is exactly what the Pentagon wants most.
There is a somewhat uncomfortable irony that needs to be clarified.
OpenAI has already practiced the "technical transparency" and "continuous monitoring" it promised to the Pentagon on its own users.
In August 2025, OpenAI quietly disclosed a new monitoring system in an official blog post about user mental health crises: when the system detects that a user is "planning to harm others," the conversation is routed to a dedicated pipeline handled by a trained human review team authorized to report the situation to law enforcement. This was voluntarily disclosed by OpenAI, but buried in the middle of a lengthy article focused on mental health, eliciting a muted response.
In February 2026, shortly before this contract was signed, OpenAI launched its advertising system and updated its privacy policy, making one thing clear: for free and basic paid users using ChatGPT, the system performs “in-session context analysis” to display relevant ads based on conversation topics. For example, if you’re discussing recipes, you might see ads for food delivery services. OpenAI emphasized that the actual content of conversations is not shared with advertisers, but the analysis occurs in real time. Ads began testing on February 9.
In November 2025, Mixpanel, OpenAI’s third-party analytics provider, was breached, exposing partial data of some API users, including names, email addresses, approximate geographic locations, operating systems, and browser information. OpenAI subsequently terminated its partnership with Mixpanel, and legal proceedings are now underway. The incident primarily affected API developer users; among regular ChatGPT users, only those who had submitted help center tickets via the platform were impacted.
This is the company that promised the Pentagon "technical transparency, continuous monitoring, and visibility into everything that happens."
What it does best is precisely to let others look in, because it has become accustomed to treating its users this way.
Anthropic believes rules can constrain users, while OpenAI believes embedding its own people is more effective than any clause. The former follows an idealistic compliance logic; the latter follows a realistic influence logic. The Pentagon chose the latter because it is more familiar and more controllable to them.
What happened during those middle hours?
It was 5:01 PM EST on February 27, corresponding to 2:00 AM Beijing Time on February 28.
Anthropic’s ultimatum deadline has passed. Dario did not compromise. Trump announced a ban on Truth Social, Hegseth posted on X labeling Anthropic as a “supply chain risk,” and Anthropic has announced it will respond through legal means.
One statement in Hegseth’s declaration is noteworthy: “Anthropic’s stance is fundamentally incompatible with American principles.” Yet, in the same statement, he said Anthropic can continue providing services to the Pentagon “for no more than six months to ensure a smooth transition.” In other words, they’ve just labeled a company a national security risk—but are still using its products. This logical contradiction has not been directly addressed by anyone.
Hours later, Altman posted that tweet.
Looking back at his earlier remarks at the all-hands meeting: he said he hoped OpenAI could “help de-escalate the situation” and find a solution that could “establish a framework for the entire industry.” This is not the tone of someone passively waiting.
This is not the first time Silicon Valley has seen this kind of operation.
In 2023, OpenAI’s nonprofit board dismissed Altman for being “not transparent enough,” citing concerns over his overly rapid pace and internal communication issues. Five days later, Altman returned with the collective support of employees, and the board was dissolved. He then led the company’s restructuring from a nonprofit to a for-profit entity, enclosing the original nonprofit mission within a new legal framework.
This time, it's called "found a common framework."
What did Dario lose?
Jerry McGinn, Director of the Center for Industrial Base at the Center for Strategic and International Studies, offered a measured assessment: “This is great PR for Anthropic, and they didn’t need that $200 million anyway.”
This assessment is financially sound. Anthropic generated $14 billion in revenue for the full year of 2025 and is valued at $380 billion; its largest shareholder, Amazon, is unlikely to reconsider its investment logic due to a single government ban. Legal avenues may offer a turning point. The IPO valuation is unlikely to suffer material damage; the narrative of “rejecting government pressure and upholding safety principles” is unlikely to be matched by any public relations budget.
But there is one thing that Dario lost.
The real-world standards for AI use in the military will be established by OpenAI inside the Pentagon, not Anthropic. That safety stack embedded in the classified network, those OpenAI researchers with security clearances, that monitoring system continuously tracking AI behavior—they will, over the coming years, grow into the de facto industry standard.
Anthropic held its ground by sticking to its principles but lost its seat in the rule-making process.
The person who took that seat is the same person who transitioned from “showing support” to “signing the contract” in less than twelve hours.
The most ironic part is that the companies taking AI safety most seriously have been pushed out of the very place where AI safety matters most.
The company that replaced it just launched a mechanism two weeks ago to integrate user conversation topics into its advertising system, experienced a third-party data breach three months ago, and earlier quietly disclosed a conversation scanning system capable of reporting users to law enforcement.
In Silicon Valley, Altman’s move within less than twelve hours has a name. It’s not called betrayal; it’s called timing.
Click to learn about the open positions at BlockBeats
Welcome to the official BlockBeats community:
Telegram subscription group: https://t.me/theblockbeats
Telegram group: https://t.me/BlockBeats_App
Official Twitter account: https://twitter.com/BlockBeatsAsia