The US government has launched a program to test artificial intelligence (AI) for potential leaks of classified data related to nuclear technology.
11 0
Photo – t4.com.ua
As NBN reports with reference to the AXIOS material, Anthropic, the company that developed Claude 3 Sonnet, and the US Department of Energy, have attracted specialists from the National Nuclear Security Administration (NNSA) to study the openness of the said AI model.
The testing process uses the “Red Teaming” technique, in which experts will deliberately try to mislead/puzzle the AI system to reveal its vulnerabilities. In this particular situation, NNSA experts intend to model scenarios in which cyberattackers attempt to extract classified information about nuclear weapons from Claude 3 Sonnet.
Anthropic added that this is the first phase of such tests for artificial intelligence (duration – until February 2024), conducted under conditions of absolute secrecy, and successful tests will help expand the scope of cooperation with government agencies, while strengthening AI security.
Earlier, we wrote about the fact that scientists showed how existing AI models coped with non-standard mathematical test.