The U.S. National Institute of Standards and Technology (NIST) conducted a significant red-teaming study on frontier artificial intelligence models just before Donald Trump's second presidential term, but the findings were never published, according to a report by Wired. This unpublished report detailed 139 vulnerabilities in advanced AI systems, including issues related to misinformation, bias, fairness, and even climate impact. The decision to withhold the report reportedly stemmed from concerns about clashing with the incoming Trump administration's evolving stance on AI regulation.
The study utilized NIST’s draft safety checklist, known as AI 600-1, which was designed to help companies assess their own AI systems. This checklist explicitly addressed areas such as misinformation and algorithmic bias. However, the incoming team for Donald Trump signaled a clear intention to remove these topics from any federal AI rules, prioritizing a focus on traditional security threats and innovation.
Donald Trump’s public “AI Action Plan” subsequently reinforced this position, mandating the framework to "strip mentions of misinformation and DEI" (Diversity, Equity, and Inclusion). This policy shift contrasted sharply with the Biden administration's previous emphasis on comprehensive AI safety, as outlined in Executive Order 14110, which aimed to manage risks and promote trustworthy AI. The Trump administration later revoked key Biden-era AI directives, including EO 14110, signaling a deregulatory approach.
Crucially, the hidden red-team report, which used the unedited AI 600-1 checklist, demonstrated that real AI models failed precisely on the issues that the new administration sought to exclude. This included generating misinformation and exhibiting biases, underscoring the practical relevance of the suppressed findings. Sources familiar with the situation, as reported by Wired, indicated that the document was one of several AI-related papers from NIST that were not released to avoid conflict with the changing political landscape.
The non-publication of this comprehensive report highlights the fragility of AI governance in a politically divided environment. Industry leaders and experts have expressed concerns that withholding such critical insights undermines efforts to standardize safety measures and could impede the development of safer AI innovation. The findings could have provided a crucial blueprint for companies to bolster their AI safety protocols, accelerating voluntary safety testing across the industry.