Author: Ada, Deep Tide TechFlow
February 24, Tuesday. Washington, Pentagon.
Dario Amodei, CEO of Anthropic, sits opposite Secretary of Defense Pete Hegseth. According to multiple media outlets including NPR and CNN, the atmosphere of the meeting was "polite," but the content was anything but courteous.
Hegseth gave him an ultimatum: lift the military use restrictions on Claude by 5:01 PM on Friday, allowing the Pentagon to use it for "all legal purposes," including autonomous weapon targeting and large-scale domestic surveillance.
Otherwise, a $200 million contract would be canceled. Activate the Defense Production Act to force requisition. Designate Anthropic as a "supply chain risk," which is equivalent to placing it on a blacklist of hostile entities like Russia and China.
On the same day, Anthropic released the third version of its "Responsible Scaling Policy" (RSP 3.0), quietly removing the core commitment since the company's inception: not training more powerful models if safety measures cannot be ensured.
On the same day, Elon Musk posted on X: "Anthropic is stealing training data on a massive scale, this is a fact." Meanwhile, X's community notes pointed out reports that Anthropic paid a $1.5 billion settlement for using pirated books to train Claude.
Within seventy-two hours, this self-proclaimed "soulful" AI company played three roles: a safety martyr, an intellectual property thief, and a traitor to the Pentagon.
Which one is real?
Perhaps all of them.
The Pentagon’s "Obey or Roll"
The first layer of the story is simple.
Anthropic is the first AI company to gain access to US Department of Defense classified information. A contract obtained last summer was capped at $200 million. OpenAI, Google, and xAI subsequently secured contracts of similar scale.
According to Al Jazeera, Claude was used in a US military operation in January of this year. The operation reportedly involved the kidnapping of Venezuelan President Maduro.
However, Anthropic drew two red lines: no support for fully autonomous weapon targeting, and no support for large-scale surveillance of US citizens. Anthropic believes that the reliability of artificial intelligence is insufficient to operate weapons, and there are currently no laws regulating the use of AI in large-scale surveillance.
The Pentagon does not buy this.
White House AI advisor David Sacks publicly accused Anthropic on X last October of "weaponizing fear and engaging in regulatory capture."
Competitors have already knelt. OpenAI, Google, and xAI all agreed to allow the military to use their AI for "all legal scenarios." Musk's Grok just received approval to enter classified systems this week.
Anthropic is the last one standing.
As of the time of publication, Anthropic stated in its latest announcement that they do not intend to compromise. But the deadline of 5:01 PM on Friday is fast approaching.
An anonymous former liaison officer between the Justice and Defense Departments expressed confusion to CNN: "How can you simultaneously designate a company as a 'supply chain risk' and force that company to work for your military?"
Good question, but this is not within the Pentagon's considerations. What they care about is that if Anthropic does not compromise, they will take coercive measures or become a Washington outcast.
"Distillation Attack": A Face-Slapping Accusation
On February 23, Anthropic published a strongly worded blog accusing three Chinese AI companies of conducting "industrial-scale distillation attacks" on Claude.
The defendants are DeepSeek, Moonshot AI, and MiniMax.
Anthropic accused them of initiating over 16 million interactions with Claude through 24,000 fake accounts, targeting the extraction of Claude’s core capabilities in agent reasoning, tool invocation, and programming.
Anthropic classified this as a national security threat, claiming that the distilled models "are unlikely to retain safety guardrails," and could potentially be used by authoritarian governments for cyberattacks, misinformation, and large-scale surveillance.
The narrative is perfect, and the timing is perfect.
Just after the Trump administration relaxed restrictions on chip exports to China, and just when Anthropic needed ammunition for its own lobbying position on chip export controls.
But Musk fired a shot: "Anthropic is stealing training data on a massive scale and has paid billions of dollars in settlements for it. This is a fact."
Tory Green, co-founder of AI infrastructure company IO.Net, stated: "You trained your model with data from the entire internet, and then when others use your public API to learn from you, that's called 'distillation attack'?"
Anthropic calls distillation an "attack," but this is routine in the AI industry. OpenAI uses it to compress GPT-4, Google uses it to optimize Gemini, and Anthropic itself is also doing it. The only difference is that this time they were the ones being distilled.
According to Erik Cambria, an AI professor at Nanyang Technological University in Singapore, stated to CNBC: "The boundaries between legitimate use and malicious exploitation are often blurred."
Ironically, Anthropic just paid a $1.5 billion settlement for using pirated books to train Claude. It trained its model using data from the entire internet and then accused others of learning from it using its public API. This isn't double standards; it's triple standards.
Anthropic intended to play the victim but ended up being exposed as the defendant.
Dismantling Safety Commitments: RSP 3.0
On the same day it faced off against the Pentagon and engaged in disputes with Silicon Valley, Anthropic released the third version of its Responsible Scaling Policy.
Anthropic Chief Scientist Jared Kaplan stated in a media interview: "We believe stopping the training of AI models does not help anyone. In the context of rapid AI development, making unilateral commitments... while competitors are advancing at full speed, is meaningless."
In other words, since others are not practicing restraint, we won't pretend anymore.
The core of RSP 1.0 and 2.0 was a hard commitment to pause training if a model's capabilities exceeded the safety measures in place. This commitment earned Anthropic a unique reputation in the AI safety circle.
But 3.0 deleted this.
Instead, it introduced a more "flexible" framework, distinguishing between safety measures Anthropic can implement and safety recommendations requiring industry-wide collaboration. A risk report will be issued every 3-6 months for external expert review.
Sounds responsible?
Chris Painter, an independent reviewer from the nonprofit organization METR, expressed after reviewing the draft of the policy: "This indicates that Anthropic believes it needs to enter 'triage mode' because the methods for assessing and mitigating risks cannot keep pace with the speed of capability growth. This further demonstrates that society is unprepared for the potentially catastrophic risks of AI."
According to TIME, Anthropic spent nearly a year discussing this rewrite internally, with CEO Amodei and the board unanimously approving it. The official explanation is that the original policy was designed to foster industry consensus, but the industry has simply not kept up. The Trump administration adopted a lax attitude toward AI development, even attempting to abolish relevant regulations in various states. Federal-level AI legislation is nowhere in sight. While establishing a global governance framework in 2023 might still be possible, three years have passed, and that door has clearly closed.
An anonymous researcher who has followed AI governance for a long time stated more directly: "RSP is Anthropic's most valuable brand asset. Deleting the pause training commitment is akin to an organic food company quietly tearing the word 'organic' off its packaging and then telling you that their testing is now more transparent."
Identity Crisis Under a $380 Billion Valuation
In early February, Anthropic completed a $30 billion funding round with a valuation of $380 billion, with Amazon as an anchor investor. Since its founding, it has achieved an annualized revenue of $14 billion. Over the past three years, this figure has increased more than tenfold each year.
At the same time, the Pentagon threatened to blacklist it. Musk publicly accused it of data theft. Its core safety commitment has been deleted. After the resignation of Anthropic's AI safety head, Mrinank Sharma, he wrote on X: "The world is in danger."
Contradictory?
Perhaps contradiction is in Anthropic's genes.
This company was founded by former OpenAI executives who were concerned that OpenAI was moving too quickly on safety issues. Then, they created their own company to build more powerful models at an even faster pace while telling the world how dangerous those models are.
The business model can be summarized in one sentence: We are more afraid of AI than anyone else, so you should pay us to create it.
This narrative operates perfectly in 2023-2024. AI safety is a hot topic in Washington, and Anthropic is the most popular lobbyist.
In 2026, the wind will shift.
"Woke AI" will become an attack label, state-level AI regulatory bills will be blocked by the White House, and while California's SB 53, supported by Anthropic, was signed into law, there will be desolation at the federal level.
Anthropic's safety narrative is shifting from a "differentiated advantage" to a "political liability."
Anthropic is performing a complex balancing act. It needs to be "safe" enough to maintain its brand while also being "flexible" enough not to be abandoned by the market and government. The problem is that the tolerances on both ends are shrinking.
How Much Is the Safety Narrative Worth Now?
Putting three pieces together, the picture becomes clear.
Accusing Chinese companies of distilling Claude is aimed at reinforcing the lobbying narrative for chip export controls. To keep up in the arms race, the safety pause commitment was deleted. Refusing the Pentagon's demand for autonomous weapons is to preserve the last layer of moral veneer.
Each step has logic, but each step contradicts the others.
You cannot claim that Chinese companies "distilling" your model poses a national security risk while simultaneously removing commitments to prevent your own model from going rogue. If the models are indeed that dangerous, you should be more cautious, not more aggressive.
Unless you are Anthropic.
In the AI industry, identity is defined not by your statements but by your balance sheet. Anthropic's "safety" narrative is essentially a brand premium.
In the early days of the AI arms race, this premium was valuable. Investors were willing to pay a higher valuation for "responsible AI," governments were willing to give the green light for "trustworthy AI," and customers were willing to pay for "safer AI."
But by 2026, this premium is evaporating.
What Anthropic now faces is not a question of "whether to compromise," but a ranking question of "who to compromise with first." Compromise with the Pentagon, and the brand is damaged. Compromise with competitors, and the safety commitments are invalidated. Compromise with investors, and both sides have to yield.
By 5:01 PM on Friday, Anthropic will deliver its answer.
But no matter what the answer is, one thing is already certain: the Anthropic that once stood on "we are different from OpenAI" is becoming just like everyone else.
The endpoint of an identity crisis is often the disappearance of that identity.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
