Collaborative Research: EAGER: Unveiling Security Risks to Scientific Integrity in AI-Driven Biological Discovery

Modern biological research increasingly relies on artificial intelligence frameworks that integrate diverse, externally sourced digital resources, including open-source datasets, reference databases, and reusable artificial intelligence models. While this openness accelerates discovery in vital sectors, such as agriculture and environmental monitoring, it may create research security vulnerabilities with potential implications for national security, scientific integrity, and the bioeconomy. Because biological AI workflows frequently incorporate external digital assets of unverified origin or quality, undetected integrity flaws or data alterations may propagate through analysis pipelines and compromise downstream scientific conclusions. Existing institutional safeguards primarily address personnel and access control, leaving a significant gap in tracking how external resource dependencies influence scientific interpretation. This project addresses this gap by identifying where these dependencies enter biological workflows, mapping their propagation, and evaluating whether standard validation and review practices can detect integrity failures before it affects scientific findings. By securing the reliability of artificial intelligence applications in the life sciences, this award advances national interests in agricultural competitiveness, environmental stewardship, national security, and the bioeconomy, while simultaneously training a cohort of graduate and undergraduate students at the intersection of cybersecurity and biology. The goals of this project are to (i) characterize how externally sourced digital resources, including datasets, annotations, metadata, software, and pretrained models, enter biological workflows; (ii) identify gaps in provenance verification; and (iii) quantify how resource integrity failures alter downstream scientific conclusions. This work is conducted in two distinct phases. First, structured surveys and interviews with practitioners across diverse biological domains serve to construct a provenance-to-interpretation graph that models the flow of external resources. A pathway-level risk scoring framework is developed to isolate high-risk dependency patterns based on provenance uncertainty, weakness of validation or oversight, and downstream biological impact. Second, empirical, adversarial integrity experiments are conducted on a soil metagenomic workflow utilizing historical agronomic data, with cross-domain validation in plant phenotyping, crop genomics, and environmental monitoring. By systematically introducing controlled perturbations into selected external artifacts, the work measures the propagation depth of integrity failures and assesses the efficacy of current validation practices. Deliverables include new methodologies for workflow exposure analysis, a graph-based representation for dependency tracking, a risk-scoring framework, and empirical evidence documenting which dependency patterns create measurable scientific-integrity concerns. 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.