Photos by Getty Images and Unsplash, design by Kekeli Sumah/RAND | |
|
|
|
What makes AGI so threatening? RAND gives Congress five answers
|
|
|
In a recent visit to Capitol Hill, RAND Vice President and Director of Global and Emerging Risks Jim Mitre briefed lawmakers on how artificial general intelligence might jeopardize U.S. national security. Testifying before the U.S. Senate Armed Services Committee's Cybersecurity Subcommittee, Mitre highlighted findings from a recently published perspective on five ways AGI might pose a national security threat. Mitre and co-author Joel Predd argue AGI could:
|
- Enable a decisive wonder weapon, akin to a nuclear weapon.
- Catalyze a systemic shift in competitive advantage.
- Empower a nonexpert to become an expert malicious actor.
- Create independent agents with enough autonomy to cause global havoc.
- Cause global instability as states complete to win an AI arms race.
|
|
|
Not all of Mitre’s testimony focused on AI’s potential downsides. Addressing a question from Sen. Jackie Rosen (D-NV) on the AI workforce, Mitre underscored the competitive edge AI could offer the U.S. military. “Those militaries that are fully embracing and applying it across a range of applications are going to be in a significant advantage relative to those militaries that aren’t,” he said. Read more: Artificial General Intelligence’s Five Hard National Security Problems.
|
|
|
|
|
Ready to respond: RAND experts deliver rapid response on AI energy supply
|
|
|
When the U.S. government seeks answers, RAND experts stand ready to deliver swift insight. One recent example shows how quickly RAND can respond to national needs. In January 2025, the National Science Foundation issued a request for information, or RFI, on how the United States can maintain AI leadership and bolster private innovation. Over a short period, RAND researchers Yusuf Mahmood, Christopher Byrd, Elika Somani, Konstantin Pilz, and Lennart Heim delivered detailed policy options. Their response focused on a looming risk: that the United States could fall behind in AI infrastructure due to energy supply constraints. To unlock energy capacity, the team proposed:
|
- Streamlining federal permitting processes.
- Ensuring reliable grid and transmission line access.
- Opening federal lands for AI-related projects.
- Ensuring infrastructure developers meet national security requirements.
|
|
|
|
|
|
|
|
|
|
|
Making it make sense: Lennart Heim breaks down new U.S. framework on AI diffusion
|
|
|
A new U.S. framework tackles a tough policy challenge: how to preserve U.S. AI leadership without stifling the global spread of beneficial AI technologies. Recognizing the framework’s complexity, RAND’s Lennart Heim set out to help U.S. policymakers make sense of it. Released just one day after the framework’s publication, Heim’s expert insight, Understanding the Artificial Intelligence Diffusion Framework, breaks down core elements. These include export controls on advanced AI chips and closed model weights, and a tiered country system for governing import restrictions, deployment limits, and other activities. Heim’s timely response garnered significant interest. In January alone, the explainer received more than 4,000 downloads — making it the most-read publication from RAND’s Technology and Security Policy Center that month. Broadening its reach, Heim also appeared on the Lawfare Daily and AI Summer podcasts, offering listeners an in-depth look at the framework’s intricacies. Read more: Understanding the Artificial Intelligence Diffusion Framework.
|
|
|
|
|
|
|
RAND experts weigh in on the DeepSeek surprise
|
|
|
In January 2025, Chinese AI company DeepSeek surprised the world by unveiling a domestically produced large language model that appeared to rival the performance of top U.S. models — seemingly at a fraction of the cost. RAND researchers and advisors were quick to offer their insights on DeepSeek and the implications of its release. In recently published articles:
|
|
|
|
|
How capable are today's AI models at chemical and biological science? Turns out: Very.
|
|
|
History shows nefarious actors have long aimed to create biological and chemical weapons — but a lack of technical expertise has stood in their way. As AI systems grow more capable, a team of RAND researchers, led by Sunishchal Dev and Charles Teague, assessed whether today’s most advanced language models possess enough biology and chemistry knowledge to assist in weapon development — and whether safety features are robust enough to prevent misuse. The findings underscore how far AI has come: when tested on scientific knowledge benchmarks, many of the 31 leading models — especially OpenAI’s o1 family — matched or outperformed human experts. But a tradeoff appeared during safety testing: removing safeguards made models more willing to assist with harmful requests, while also diminishing their performance on some knowledge tasks. To address the growing capabilities of today’s AI models, the authors call for stronger human baselines, more-challenging benchmarks, and better standardization for evaluating risky model behavior. Read more: Toward Comprehensive Benchmarking of the Biological Knowledge of Frontier Large Language Models.
|
|
|
|
|
|