Anthropic is collaborating with the US Department of Energy (DOE) to test its Claude 3 Sonnet model. The experiment aims to evaluate the AI’s ability to “not share” sensitive or potentially dangerous information related to nuclear energy, particularly the creation of weapons.
Since April 2024, specialists from the DOE’s National Nuclear Security Administration (NNSA) have been conducting rigorous assessments of Claude 3 Sonnet. These tests, often referred to as “red requests,” involve attempts to manipulate the AI system to uncover vulnerabilities that could potentially be exploited to create nuclear weapons.
Innovative Approach to National Security
Anthropic asserts that this initiative is the first of its kind, conducted within a highly secure environment. According to Marina Favaro, Anthropic’s director of national security policy, the project could pave the way for similar collaborations with other government agencies. “While U.S. industry leads in AI development, the federal government is building unique expertise in assessing AI for national security risks,” Favaro stated.
The NNSA also highlighted the significance of this collaboration. Wendin Smith, NNSA’s deputy administrator, emphasized that artificial intelligence is “one of the key technologies requiring sustained attention for national security purposes.”
Future Prospects and Ongoing Updates
The pilot program, which includes testing the newer Claude 3.5 Sonnet, will run through February, notes NIXsolutions. Anthropic plans to continue its partnership with the government, focusing on enhancing safeguards and ensuring its systems meet stringent security standards. The company has committed to sharing testing results with academic labs and other organizations to further AI safety research.
This collaboration represents a critical step in using AI responsibly for national security purposes. As the program progresses, we’ll keep you updated on the outcomes and any new developments in the field.