Anthropic Weakens AI Safety Commitments Amid Pentagon Ultimatum Over Military Use

Safety Takes a Back Seat to Competition

Anthropic — long regarded as one of the most safety-conscious AI labs — has abandoned a foundational commitment in its own safety framework. The company cited the need to remain competitive in a rapidly advancing market as the driving reason behind the change.

In 2023, Anthropic introduced its Responsible Scaling Policy (RSP), a set of voluntary rules designed to mitigate catastrophic risks in AI development. The original document committed the company to pausing AI development if its models were deemed potentially dangerous.

Anthropic's chief scientist Jared Kaplan said the company concluded that simply stopping model training would not help anyone. Given the rapid pace of progress across the industry, he argued, it would be unwise to take on unilateral commitments while competitors continue to push forward.

The third version of RSP codifies a new approach: Anthropic will continue improving its AI as long as it believes its models do not hold a significant advantage over competitors. In a blog post, the company noted that the political environment has shifted toward prioritizing competitiveness and economic growth, while AI safety discussions have failed to gain meaningful support at the federal level.

The updated policy does include greater transparency measures, such as detailed publication of model testing results. Anthropic says it will only slow development if it becomes the undisputed leader in the AI race while simultaneously identifying a significant catastrophe risk.

When RSP launched in 2023, the company hoped rivals would adopt similar pledges. None did — no other major player made a comparable commitment to halt AI development upon detecting risks.

Why This Matters

Anthropic's retreat from its own safety framework marks a troubling precedent. Even companies that built their brand on responsible AI development are now backing away from voluntary commitments under competitive pressure and a lack of regulatory enforcement. The failure of self-regulation raises fundamental questions about whether the AI industry can govern itself without binding legislation.

The Pentagon's Hard Deadline

Simultaneously, Anthropic is locked in a standoff with the U.S. Department of Defense. The conflict centers on the Pentagon's plans to use AI for mass surveillance of American citizens and autonomous weapons development — applications Anthropic explicitly opposes.

Pentagon officials have stated their intent to use large language models "for all lawful scenarios" without restrictions and floated the possibility of terminating their contract with the company.

Anthropic CEO Dario Amodei met with Defense Secretary Pete Hegseth to discuss the situation. The department issued an ultimatum: the company must accept government terms by February 27. If Anthropic refuses, authorities could designate it as a supply chain threat, jeopardizing the startup's business with other U.S. government contractors. The alternative is invoking the Defense Production Act, which would allow the Pentagon to forcibly access Anthropic's technology.

Franklin Turner, a government contracts attorney at McCarter & English, called the scenario unprecedented, noting it would almost certainly trigger a wave of litigation if the administration takes adverse action against Anthropic.

In response, Anthropic stated that both parties "continue to engage in good-faith dialogue."

Anthropic Holds the Line on Military Use

Despite the pressure, Reuters reports that Anthropic has no plans to soften its restrictions on military applications of its technology. The company continues to maintain two red lines: no fully autonomous weapons and no mass surveillance of Americans.

Ethereum co-founder Vitalik Buterin voiced his support for the company's stance:

Buterin said his opinion of the company would significantly improve if Anthropic refuses to back down and honorably accepts the consequences.

In February, it emerged that the Claude model was used in an operation to capture Venezuelan President Nicolás Maduro. According to reports, Amodei did not express concern about this use of the company's products during his meeting with Hegseth.

The Economist notes that the severity of the Pentagon's ultimatum signals the military's reluctance to abandon Claude — suggesting the startup's technology may be irreplaceable for defense applications.

Separately, SpaceX and its AI subsidiary xAI are participating in a classified Pentagon tender to develop voice-controlled autonomous drone swarms.