CICI: IPAAI: VERITAS - VERified Infrastructure for Trustworthy AI in Science

Artificial intelligence (AI) has become essential to scientific discovery, yet its rapid adoption introduces risks that traditional cybersecurity controls cannot detect. Conventional security tools cannot tell whether an AI model has been tampered with so that it runs normally but still produces wrong scientific answers. The objective of VERITAS is to establish AI Assurance as a function that is distinct from, but complementary to, conventional cybersecurity. The project addresses this through three coordinated tasks. First, through introduction of standardized documentation of model cards and dataset datasheets for substantial GPU allocations on national computing resources. Second, through deployment of an AI Assurance Engineer to inspect AI projects for hidden weaknesses, supported by an expert "red team" which safely probes systems the way an attacker might. Finally, through hands-on challenges through the National Data Platform Education Hub that train students nationwide to spot these problems themselves. The results will include validated AI workflows, a catalog of scientific-AI attack patterns and mitigations, and a trained AI-security workforce. By building these protections into the research infrastructure scientists already use, VERITAS makes AI-driven science more trustworthy and reliable for the nation. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.