US Gets Pre-Release Access to AI Models from Five Labs for Security Tests
The Center for AI Standards and Innovation, a unit within the US Department of Commerce, now holds pre-release access to advanced AI models from five leading laboratories. This setup allows national security testing before public availability. The agency recently finalized agreements with Google DeepMind, Microsoft, and xAI.
New Agreements Expand Testing Capabilities
These fresh pacts enable CAISI to examine the models for potential risks to national security. AI developers share early versions, including some with lowered safety protections, to support thorough assessments. The center has conducted over 40 evaluations already, with several involving models not yet released.
CAISI Director Chris Fall stated, "Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications." Such testing occurs in controlled settings, including classified ones, to identify vulnerabilities.
Building on Prior Partnerships
The latest deals extend previous arrangements with Anthropic and OpenAI. Those earlier agreements focused on collaborative safety reviews and studies on reducing risks. Anthropic, known for its Claude AI series, emphasizes constitutional AI principles in development. OpenAI, creator of the GPT models, has long partnered with US entities on safety.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Google DeepMind, a Google subsidiary, leads in areas like AlphaFold for protein prediction and Gemini models. Microsoft integrates AI deeply into Azure cloud services and partners with OpenAI. xAI, founded by Elon Musk in 2023, aims to understand the universe through models like Grok.
Context of Rapid AI Advances
This expansion arrives amid fast progress in AI capabilities. Models now excel at detecting and using security flaws more effectively. Competition with China in technology adds urgency to these measures.
CAISI, established under the National Institute of Standards and Technology within Commerce, sets benchmarks for AI reliability. Frontier AI refers to the most powerful systems pushing computational limits. Labs provide these for preemptive checks to mitigate threats like unauthorized access or misuse.
Joint efforts include research on safeguards. Classified testing ensures sensitive data handling. Over 40 evaluations demonstrate active progress, covering diverse risks from model behaviors.
The five labs represent key players: OpenAI transformed chatbots with GPT-3 in 2020; Anthropic spun out from OpenAI in 2021; DeepMind joined Google in 2014; Microsoft invests billions in AI; xAI launched swiftly with backing from Musk's network.
These steps reflect government commitment to balancing innovation and security as AI integrates into critical sectors.