US and tech firms strike deal to review AI models for national security before public release

The US government, through the Center for AI Standards and Innovation (CAISI), has finalized agreements with major tech firms including Google DeepMind, Microsoft, and xAI. These deals will allow CAISI to review early versions of powerful new AI models before their public release. The collaborations aim to understand the capabilities of these advanced AI systems and mitigate potential national security risks, particularly concerning cybersecurity, biosecurity, and chemical weapons. This initiative builds on similar agreements made with OpenAI and Anthropic two years ago, with CAISI having already conducted over 40 evaluations. The reviews involve examining models with reduced safety guardrails to thoroughly assess national security implications, addressing growing concerns about the potential misuse of advanced AI.

Microsoft, Google, and xAI have agreed to grant the U.S. federal government access to their advanced artificial intelligence models for national security testing. This initiative, announced by the Department of Commerce's Center for AI Standards and Innovation (CAISI), allows U.S. officials to evaluate these AI systems before deployment, assessing their capabilities and potential security risks. The agreement aims to identify threats such as cyberattacks and military misuse posed by powerful AI, especially in light of recent concerns about models like Anthropic's Mythos. This move fulfills a previous administration pledge to partner with tech companies on vetting AI for national security. Microsoft will collaborate with U.S. government scientists on testing and developing shared datasets for evaluating AI systems, mirroring a similar agreement with the UK's AI Security Institute.

The US Department of Defense has announced a new agreement with seven AI companies, including SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and Amazon Web Services, to integrate their advanced AI technologies into classified systems. This move aims to accelerate the military's transformation into an AI-first fighting force, enhancing decision superiority for warfighters. The capabilities will be utilized in secure information systems to improve data synthesis, situational understanding, and decision-making in complex environments. Notably absent from the agreement is Anthropic, following a public dispute over access to its AI programs and concerns about their potential use in surveillance and autonomous weapons. The Pentagon previously labeled Anthropic a "supply chain risk" after the company resisted providing unrestricted access to its Claude AI.

Google has reportedly signed a classified artificial intelligence deal with the US Pentagon, allowing the military to use its AI models for "any lawful government purpose." This agreement places Google alongside other tech firms like OpenAI and xAI that are supplying AI for classified military operations. The Pentagon aims to leverage these AI tools for sensitive work such as mission planning and weapons targeting, with contracts potentially worth up to $200 million each. While the deal includes stipulations against domestic mass surveillance and autonomous weapons without human oversight, it does not grant Google veto power over lawful government decisions. Google stated its commitment to responsible AI use in national security, supporting agencies on both classified and non-classified projects.