What’s behind Anthropic’s Pentagon standoff, and its AI safety U-turn?

US-based artificial intelligence firm Anthropic has overhauled its core AI safety framework, removing a key self-imposed commitment to pause development of more powerful models and replacing it with a more flexible, nonbinding structure.

In a blog post this week, Anthropic released Version 3.0 of its Responsible Scaling Policy (RSP), acknowledging that elements of its two-year-old framework could hinder its ability to compete in a rapidly evolving AI market.

Anthropic was founded by former OpenAI executives who had voiced concerns about the risks posed by increasingly capable AI systems. The company has often positioned itself as a safety-first developer and previously described its approach as building AI with a “soul”.

Under its earlier policy, Anthropic had committed to pausing the training of more capable models if their abilities outstripped the company’s capacity to ensure they were safe and controllable.

The company argued that unilateral pauses by “responsible” developers, while competitors press ahead, could “result in a world that is less safe.”

“The developers with the weakest protections would set the pace, and responsible developers would lose their ability to do safety research and advance the public Benefit,” the policy document stated.

Instead of binding guardrails tied to preset capability thresholds, Anthropic said it will now separate its own safety plans from the industry-wide recommendations.

The revised framework includes a “Frontier Safety Roadmap” outlining public goals across security, safeguards, alignment and policy.

“These are not hard commitments but rather public goals against which we will openly grade our progress,” the company said.

Anthropic vs Pentagon

The announcement comes the same week the company is locked in a high-stakes dispute with the US Department of Defense.

In July, the Department of Defense awarded Anthropic a USD 200 million contract to develop AI capabilities aimed at advancing US national security.

Anthropic is reportedly the only AI company whose model, Claude, is currently deployed on the Pentagon’s classified networks.

A senior Pentagon official told US media that Grok, owned by xAI, is willing to operate in classified settings, and other AI companies are close behind.

Last month, the Pentagon signalled it intends to accelerate AI adoption, saying the technology could help the military “rapidly convert intelligence data” and make warfighters “more lethal and efficient.”

The confrontation reportedly intensified after the US military used Claude during an operation in January to capture former Venezuelan President Nicolás Maduro, according to media reports. Anthropic has said it did not discuss the use of Claude for specific operations with the Department of War.

Anthropic has reportedly sought assurances that its AI system would not be used for mass surveillance of Americans and would not make final targeting decisions in military operations without human involvement.

The company maintains that Claude, like other large language models, is susceptible to hallucinations and is not reliable enough to avoid potentially lethal errors without human oversight.

Pentagon officials have rejected suggestions that the issue involves unlawful surveillance or autonomous weapons.

In a January speech, Defense Secretary Pete Hegseth declared that the Pentagon would not use AI models that constrain lawful military operations, adding that “Department of War AI will not be woke. It will work for us.”

According to sources cited in US media, Hegseth has given Anthropic until Friday to provide the US military unrestricted use of its AI technology or risk losing its USD 200 million contract and being effectively blacklisted from future government work.

Pentagon officials are reportedly considering invoking the Defense Production Act on national security grounds or designating Anthropic a “supply chain risk” if no agreement is reached.

In its blog post, the company acknowledged that its earlier policy was designed in part to encourage a “race to the top”, pushing other AI developers to adopt stronger safeguards.

It conceded that this expectation has not fully materialised and that the anti-regulatory political climate in Washington has shifted towards prioritising AI competitiveness and economic growth.

Anthropic noted that “safety-oriented discussions have yet to gain meaningful traction at the federal level.”

The revised framework introduces periodic Risk Reports and, in certain cases, external reviews by independent AI safety experts. The company said it would publicly detail gaps between its own safeguards and the measures it believes the industry should adopt.

Anthropic has previously published research highlighting potential risks in advanced AI systems, including scenarios where models could engage in manipulative behaviour under specific conditions. It has also supported advocacy efforts around AI safeguards.