Anthropic has changed its own security rules so that the company no longer promises to automatically halt development of new Claude models unless all security requirements are met in advance.
Anthropic has changed its own security rules so that the company no longer promises to automatically halt development of new Claude models unless all security requirements are met in advance.
According to Engadget, the company has updated its Responsible Scaling Policy. This is a document in which Anthropic describes when and under what conditions it should «press pause» when building more powerful models. Previously, the central promise was that if Anthropic could not guarantee specific safety measures in advance, training a new model should be stopped. Now the approach is different — without a hard rule «stop under any conditions», the company offers a softer system of controls.
In the new version of the policy, Anthropic is relying on public risk reports and security roadmaps. That is, the company promises to regularly explain what threats it sees in its models, what restrictions and checks it is adding, and what it plans to change next. Anthropic explains its decision with competition: they say that if it stops alone and other players continue to release models without sufficient safeguards, this will not make the world safer.
The changes coincided with reports of pressure from the Pentagon. According to the report, US Defense Secretary Pete Hegseth has demanded that Anthropic expand military access to Claude and loosen restrictions on its use, threatening sanctions and leverage such as defense procedures that could affect the company’s contracts and partners. At the same time, Anthropic, according to Engadget, does not want to agree to scenarios of mass surveillance of Americans or the use of AI in weapons that can operate without human decision.
The article also notes that Claude is already being used in the most sensitive US defense tasks, so the conflict around «security restrictions» has become not theoretical for Anthropic, but practical — between state requirements, market competition, and its own rules of responsible use.
Earlier, dev.ua wrote about how US Secretary of War Pete Hegseth made a tough demand to the CEO of AI company Anthropic, Dario Amodei. The Pentagon demands that the military be given unhindered access to the advanced artificial intelligence model Claude without any restrictions by Friday evening.
