The results of a unique study conducted in the US to assess the security aspects of artificial intelligence have not been disclosed to the public.
It is reported that the results of the first large-scale test of artificial intelligence systems' “red teaming” – i.e., pressure tests – organized by the US National Institute of Standards and Technology (NIST) in October last year, despite being completed before the start of the second term of the Trump administration, have not been published.
At this event held in Arlington, attended by over 30 researchers, advanced AI systems such as Meta's LLaMA model, Synthesia's artificial avatar platform, and Robust Intelligence's defense system were tested in scenarios such as leaking data, spreading false information, and creating cyberattacks. As a result, 139 new vulnerability methods were discovered.
The event was conducted in accordance with NIST's AI 600-1 Risk Management Framework. However, participants stated that some categories of this framework are vague and difficult to apply in practice.
So why wasn't it published?
According to sources who spoke to the “Wired” publication:
- The report was not published to avoid conflict with the Trump administration.
- Several NIST AI documents were shelved because they touched on topics such as DEI (Diversity, Equity, and Inclusion), disinformation, and climate change.
- The Trump administration was preparing to revoke Biden's executive orders on artificial intelligence and remove DEI-related references from AI policies.
- Interestingly, Trump's AI Action Plan specifically recommended conducting such red-teaming exercises.
Real threats
During the research:
- Multilingual questions that prompted the LLaMA model to provide information on terrorism were formulated (in Russian, Gujarati, Marathi, and Telugu).
- Serious gaps were discovered in data leakage and user data protection.
- Risks of AI systems creating emotional attachments and being susceptible to manipulation were identified.
What do scientists say?
Eliz Chien Jang, a doctoral student from Carnegie Mellon University, stated:
“If this report had been published, it would have been clear how the NIST framework works in the context of red teaming and where it falls short.”
Other participants also believe that useful results had been obtained for the scientific community, but political considerations took precedence.
The non-publication of the report indicates a clash between scientifically-based regulation and political interests in the field of artificial intelligence security. Experts believe that concealing such important documents could undermine US leadership in global AI security discussions.
Farid Alizade