@diabatem
AI Safety & Biosecurity Researcher | Frontier Model Evaluation, Red-Teaming & Dangerous Capability Assessment | Biomedical Sciences
https://www.linkedin.com/in/mariame-diabate/$0 in pending offers
I am a researcher primarily interested in frontier biological risk. I design high-severity evaluations that expose AI capabilities and failure modes in dual-use biological domains, and I work on whether those evaluations stay valid under adversarial conditions.
► Develop high-severity biological-risk evaluations spanning dual-use research of concern, gain-of-function research, pathogen transmissibility and virulence, select-agent scenarios, biosafety and containment, DNA synthesis screening, and pandemic-relevant research
► Design adversarial trajectories exceeding 100 turns to test whether autonomous safety monitors detect escalating malicious use when risk emerges cumulatively across extended context rather than from a single overtly malicious prompt
► Led a six-person master's/PhD team on SPEC-GAP, a study of indirect prompt-injection propagation across 216 trajectories, showing that a third delegation hop cut unauthorized execution from 66.7% to 22.2% while execution-or-propagation held at 90.7%, and that injection information remained detectable at an executor that never read the poisoned document (AUROC 0.943)
► Developed PhD-level biology evaluation tasks exposing frontier-model reasoning weaknesses on complex scientific problems requiring multi-step inference, comparing responses across six model versions with expert preference judgments
Nothing yet.