Pentagon reportedly threatens to cut off Anthropic

- The core of the dispute centers on Anthropic's refusal to remove safeguards against its models being used for fully autonomous weapons and mass domestic surveillance. The Pentagon is pushing for an "all lawful purposes" standard, which would cover sensitive areas like weapons development and intelligence collection. - Other major AI labs, including OpenAI, Google, and xAI, have reportedly been more flexible in negotiations with the Pentagon. These companies have apparently agreed to remove standard guardrails for their work with the Department of Defense. - This conflict is not just theoretical; a report from The Wall Street Journal alleged that Anthropic's Claude model was used via a partnership with Palantir in a US military operation to capture former Venezuelan President Nicolás Maduro. Anthropic's usage policies explicitly prohibit using its models to facilitate violence or develop weapons. - The disagreement puts a contract potentially worth up to $200 million at risk for Anthropic and could lead to its replacement with other AI providers on classified networks. - Anthropic's stance is rooted in its "Constitutional AI" approach, a method for training models to be harmless by giving them a set of principles to follow, rather than relying solely on human feedback. This technical framework is central to its brand of developing safe and steerable AI. - The Department of Defense formally adopted five ethical principles for AI in 2020: Responsible, Equitable, Traceable, Reliable, and Governable. The "Governable" principle includes the ability to "disengage or deactivate deployed systems that demonstrate unintended behavior." - The Pentagon's push for unrestricted access is part of a broader effort to integrate AI across all classification levels for tasks like mission planning and weapons targeting. Officials have expressed frustration with company-imposed restrictions, arguing that as long as the use complies with US law, it should be permitted.