A suppressed NIST study reveals 139 vulnerabilities uncovered during a red-teaming exercise on leading AI systems, including Meta's Llama. The unpublished report also identified flaws in the government's AI risk assessment framework, withheld amid political pressure during the presidential transition.