Using variants of cystine-dense peptides (CDPs) from venoms and other natural sources, we established a half-billion compound mini-protein library for drug discovery. Using a combination of Rosetta and I-TASSER, we predicted the structure of each CDP which enabled in silico docking simulations of library CDPs to targets of interest. After narrowing candidates to dozens, we expanded each by methionine and tyrosine substitutions (to create hydrophobic patches) and used surface display screening to identify a low nanomolar binder to PD-L1. We used this binder to engineer a PD-L1CDP:CD3 bispecific T cell engager (BTE) that outperformed an antibody-based bispecific in multiple in vitro and in vivo models. We since assessed activity in multiple models of diffuse midline glioma (DMG, aka DIPG), a uniformly fatal pediatric brain tumor. In vivo, the BTE plus activated T cells cured 90% of the mice (no evidence of disease at 120 days by histology) whereas of the control mice succumbed to cancer within 18 days. We next engineered a B7-H3 CAR T cell that secretes the PD-L1CDP:CD3. Because interferon gamma, released by T cells engaged in killing cancer cells, causes upregulation of PD-L1 on cancer cells, the CAR T and secreted BTE were able to kill both B7-H3-positive and B7-H3-negative cancer cells. This strategy demonstrated a unique capability of addressing both tumor heterogeneity and upregulation of PD-L1 which are normally both mechanisms of CAR T cell resistance. In this case, the protective mechanism is converted into a therapeutic vulnerability.