Anthropic Chief Executive Declines Pentagon Request to Remove Artificial Intelligence Safety Protocols

Anthropic, the artificial intelligence company behind the Claude language model, has refused to comply with demands from the United States Department of Defense to eliminate safety restrictions governing the deployment of its technology for sensitive military applications.

Chief Executive Dario Amodei disclosed that the Pentagon specifically sought the removal of safeguards preventing the utilisation of Claude for mass domestic surveillance operations and fully autonomous weapons systems. Amodei confirmed these applications have never been incorporated within existing contractual arrangements with the Department of War, the alternative designation for the Defense Department established through executive order.

The company maintained its position despite facing potential regulatory consequences, with Amodei stating that should the Department opt to terminate Anthropic’s services, the organisation would facilitate an orderly transition to alternative providers. An Anthropic representative elaborated on Thursday that revised contractual language presented by the DoD represented minimal progress towards preventing Claude’s deployment for mass surveillance of American citizens or fully autonomous weaponry.

The representative characterised newly proposed safeguards as containing legal provisions permitting their circumvention at the Department’s discretion, noting these restricted protections have constituted the central focus of negotiations extending over several months.

Emil Michael, United States Undersecretary for Defense, issued pointed criticism of Amodei on social media platform X, alleging the executive sought to exercise personal control over military operations whilst accepting risks to national security. Pentagon officials previously indicated that non-compliance could result in invocation of the Defense Production Act against Anthropic, legislation granting presidential authority to mandate corporate fulfilment of defence requirements.

Additional measures under consideration include designating Anthropic as a supply chain security risk, effectively precluding government utilisation of its systems. A former Department of Defense official characterised the legal foundations for such actions as substantively weak.

Sources familiar with negotiations disclosed that tensions between Anthropic and the Pentagon predate public revelations regarding Claude’s involvement in a United States operation targeting Venezuelan President Nicolás Maduro. Whilst Amodei did not delineate specific mechanisms through which the technology could enable mass surveillance or autonomous weapons deployment, he noted in a corporate communication that artificial intelligence systems possess capabilities to aggregate disparate data into comprehensive individual profiles at significant scale.

Amodei distinguished between legitimate foreign intelligence gathering, which the company supports, and domestic mass surveillance, which he characterised as incompatible with democratic principles. Regarding weaponised applications, Amodei assessed contemporary AI systems as insufficiently reliable to power fully autonomous weapons, citing risks to military personnel and civilian populations absent adequate oversight mechanisms.

The executive emphasised that existing guardrails remain inadequate for such deployments, proposing collaborative research and development arrangements with the Department of War to enhance system reliability. The Pentagon has declined this proposal to date.