Anthropic and Pentagon Clash Over AI Safeguards [Model Behavior]

[00:00] Nina Park: Welcome to Model Behavior. We examine how AI systems are built, deployed, and operated [00:07] Nina Park: in real professional environments. Joining me today is our correspondent, Thatcher. [00:13] Nina Park: Thanks, Nina. Today we start with a significant infrastructure expansion for Google Cloud. [00:18] Nina Park: Google and Liberty Global have announced a five-year strategic partnership that puts Gemini AI models at the center of the European Telecom Operators Digital Transformation. [00:30] Nina Park: The deal covers approximately 80 million fixed and mobile connections, including Virgin Media O2 in the UK and Telanet in Belgium. [00:39] Nina Park: Satcher, it's notable that the integration spans both customer-facing products and internal network operations. [00:47] Nina Park: This follows a broader trend of hyperscalers moving deeper into the telecom stack. [00:52] Nina Park: However, we're also seeing a shift in the relationship between major providers and their model partners. [00:58] Nina Park: Microsoft AI chief Mustafa Salomon confirmed recently that Microsoft is pursuing true self-sufficiency by developing internal models to reduce its dependence on open AI. [01:09] Nina Park: Right. That shift towards self-sufficiency is a critical strategic move for Microsoft as they eye the enterprise market. [01:17] Nina Park: While they continue to offer open AI-powered features, we're seeing the massive scale of [01:22] Nina Park: their current deployments. [01:24] Nina Park: For example, the Australian bank Westpac recently rolled out Microsoft 365 co-pilot to its [01:30] Nina Park: entire global workforce of 35,000 people. [01:34] Nina Park: Nina, this is currently one of the largest corporate AI assistant rollouts to date. [01:39] Nina Park: It certainly demonstrates the reach Microsoft currently holds. [01:43] Nina Park: But our lead story involves a growing rift between the public sector and AI safety-focused labs. [01:51] Nina Park: Joining us today is Chad Thompson, who brings a systems-level perspective on AI, automation, and security. [01:58] Nina Park: Chad, what's driving the current tension between Anthropic and the Pentagon? [02:04] Thatcher Collins: Nina, it centers on usage policies versus operational utility. [02:09] Thatcher Collins: Reports today indicate the Pentagon may cut ties with Anthropic, potentially avoiding a $200 million contract. [02:16] Thatcher Collins: The friction stems from the revelation that Claude was used in the capture of Nicolas Maduro in Venezuela. [02:23] Thatcher Collins: Anthropic CEO Dario Amo Dei has been vocal about restricting AI from lethal operations or mass surveillance. [02:31] Thatcher Collins: But the Defense Department is demanding models they can use for all lawful warfighting purposes. [02:37] Chad Thompson: The conflict is quite public. [02:39] Chad Thompson: Defense Secretary Pete Hegsath recently noted that the agency will prioritize models that don't restrict how the military fights wars. [02:49] Chad Thompson: Chad, it seems the Pentagon is already looking toward alternatives like XAI or Palantir's [02:55] Chad Thompson: integrated solutions, if Anthropic maintains these hard lines on usage. [03:00] Nina Park: Amadei recently argued in an essay that democracies should use AI for defense in ways that do [03:08] Nina Park: not mirror autocratic adversaries. [03:11] Nina Park: This rift underscores the challenge for AI labs trying to balance commercial safety missions [03:19] Nina Park: with the high-stakes requirements of national security contracts. [03:24] Nina Park: Thatcher, it appears this will be a defining debate for the 2026 defense budget. [03:31] Nina Park: Thank you for listening to Model Behavior, a neural newscast editorial segment. [03:37] Nina Park: For more technical details on these stories, visit mb.neuralnewscast.com. [03:43] Nina Park: Neural newscast is AI-assisted human-reviewed. [03:48] Nina Park: View our AI transparency policy at neuralnewscast.com.