Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Open Weight Risks topic
No spam. Unsubscribe anytime.
Researchers warn open-weight AI models pose unique risks to critical infrastructure and national security
Summary
Lawrence Livermore and other witnesses told the Assembly that open-weight models—whose parameters are downloadable and unmonitored—create elevated risks for national-security systems, and that many leading open-weight models currently originate from Chinese labs.
A Lawrence Livermore National Laboratory official told the Assembly that open-weight AI models raise distinct national-security concerns because they can be downloaded, modified and run locally without ongoing monitoring.
"Once downloaded, these models run locally. No API call, no usage monitoring, no terms of service enforcement," said Dr. Nate Gleason, program leader for cyber and infrastructure resilience at Lawrence Livermore. He explained that open-weight models allow operators to remove or alter any guardrails after distribution and noted that, currently, many leading open-weight models are produced by Chinese labs. "That is not an acceptable long term position for national security," Gleason said.
Gleason described Department of Energy-funded testbeds (AI FORTS) that will measure adversarial risk, assess models' resistance to manipulation, and evaluate performance on operational technology use cases. He argued that the state and federal governments should invest to ensure domestic, trustworthy open-weight options exist for air-gapped critical systems and emphasized adversarial testing and model verification for OT networks.
