Background: Cytokines orchestrate antimicrobial immunity through cell-surface receptors, and pathogens subvert this by encoding mimics and decoys that hijack the same druggable interfaces. Yet cytokine therapeutics built mostly from human and mouse data ignore the millions of years of evolution that shaped these proteins, which genome sequencing and deep learning can now exploit. Methods: We built a curated 533-protein immune-signalling panel and a vertebrate ortholog atlas across ~4,100 genomes, with confounder-controlled per-residue evolutionary features (selection, coevolution, ancestral reconstruction, structure). Protein Language Models (PLMs) supply an implicit expectation of tolerated variation, complemented by the vertebrate evolutionary record. Both, alongside structure, are integrated to predict pathogenicity, druggable surfaces, and engineering priorities with antagonised interfaces validated structurally and energetically. Results: Selection is organised with buried residues most constrained and surface residues least constrained, and by function with host-defence proteins (defensins, interferons, chemokines) fastest and intracellular signalling slowest evolving. Against arms-race expectations, antagonised interfaces evolve no faster than comparable surface and are often conserved; the druggable targets are host-tolerant sites. Independent viruses evolve unrelated decoys converging on a cytokine's obligate residues, whereas ligand mimics stay on the native footprint exposing virus-specific sites, as at the cytomegalovirus IL-10 interface, druggable to block one virus while sparing host signalling. Explicit evolutionary analysis adds information beyond what PLMs capture, greatest for under-represented, rapidly-evolving families where standard models are weakest. Conclusion: Our evolution-to-AI framework yields virus-selective targets, cross-viral decoy and host-tolerant mimic sites, generalising across pathogens, a route to AI-enabled therapy.